Digital Photogrammetry - TEC-Digital
Digital Photogrammetry - TEC-Digital
Digital Photogrammetry - TEC-Digital
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
<strong>Digital</strong> <strong>Photogrammetry</strong><br />
Architectural photogrammetry i<br />
<strong>Photogrammetry</strong>, the use of photography for surveying, primarily facilitates the<br />
production of maps and geographic databases from aerial photographs. Along with<br />
remote sensing, it represents the principal means of generating data for geographic<br />
information systems.<br />
<strong>Photogrammetry</strong> has undergone a remarkable evolution in recent years with its<br />
transformation into ‘digital photogrammetry’. First, the distinctions between<br />
photogrammetry, remote sensing, geodesy and GIS are fast disappearing, as data<br />
can now be carried digitally from the plane to the GIS end-user. And second, the<br />
benefits of digital photogrammetric workstations have increased dramatically. The<br />
comprehensive use of digital tools, and the automation of the processes, have significantly<br />
cut costs and reduced processing time. The first digital aerial cameras have<br />
become available, and the introduction of more and more new digital tools allows<br />
the work of operators to be simplified, without the same need for stereoscopic<br />
skills. Engineers and technicians in other fields are now able to carry out photogrammetric<br />
work without reliance on specialist photogrammetrists.<br />
This book shows non-experts what digital photogrammetry is and what the software<br />
does, and provides them with sufficient expertise to use it. It also gives<br />
specialists an overview of these totally digital processes from A to Z. It serves as<br />
a textbook for graduate students, young engineers and university lecturers to<br />
complement a modern lecture course on photogrammetry.<br />
Michel Kasser and Yves Egels have synthesized the contributions of 21 topranking<br />
specialists in digital photogrammetry, lecturers, researchers, and production<br />
engineers, and produced a very up-to-date text and guide.<br />
Michel Kasser, who graduated from the École Polytechnique de Paris and from<br />
the École Nationale des Sciences Géographiques (ENSG), was formerly Head of<br />
ESGT, the main technical university for surveyors in France. A specialist in instrumentation<br />
and space imagery, he is now University Professor and Head of the<br />
Geodetic Department at IGN-France.<br />
Yves Egels, who graduated from the École Polytechnique de Paris and from ENSG,<br />
is senior photogrammetrist at IGN-France, where he pioneered analytical plotter<br />
software in the 1980s and digital workstations on PCs in the 1990s. He is now<br />
Head of the Photogrammetric Department at ENSG and lectures in photogrammetry<br />
at various French technical universities.
ii Pierre Grussenmeyer, Klaus Hanke, André Streilein
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
<strong>Digital</strong> <strong>Photogrammetry</strong><br />
Michel Kasser and Yves Egels<br />
London and New York<br />
Architectural photogrammetry iii
First published 2002<br />
by Taylor & Francis<br />
11 New Fetter Lane, London EC4P 4EE<br />
Simultaneously published in the USA and Canada<br />
by Taylor & Francis Inc,<br />
29 West 35th Street, New York, NY 10001<br />
Taylor & Francis is an imprint of the Taylor & Francis Group<br />
This edition published in the Taylor & Francis e-Library, 2004.<br />
© 2002 Michel Kasser and Yves Egels<br />
All rights reserved. No part of this book may be reprinted or<br />
reproduced or utilised in any form or by any electronic, mechanical,<br />
or other means, now known or hereafter invented, including<br />
photocopying and recording, or in any information storage or<br />
retrieval system, without permission in writing from the publishers.<br />
Every effort has been made to ensure that the advice and<br />
information in this book is true and accurate at the time of going to<br />
press. However, neither the publisher nor the authors can accept<br />
any legal responsibility or liability for any errors or omissions that<br />
may be made. In the case of drug administration, any medical<br />
procedure or the use of technical equipment mentioned within this<br />
book, you are strongly advised to consult the manufacturer’s<br />
guidelines.<br />
British Library Cataloguing in Publication Data<br />
A catalogue record for this book is available from the<br />
British Library<br />
Library of Congress Cataloging in Publication Data<br />
Kasser, Michel<br />
<strong>Digital</strong> photogrammetry/Michel Kasser and Yves Egels.<br />
p. cm<br />
Includes bibliographical references and index.<br />
I. Aerial photogrammetry. 2. Image processing – <strong>Digital</strong> techniques.<br />
I. Kasser, Michel. II. Title.<br />
TA 593.E34 2001<br />
526.9′82 – dc21 2001027205<br />
ISBN 0-203-30595-7 Master e-book ISBN<br />
ISBN 0-203-34383-2 (Adobe eReader Format)<br />
ISBN 0–748–40945–9 (pbk)<br />
ISBN 0–748–40944–0 (hbk)
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Contents<br />
List of colour plates viii<br />
List of contributors ix<br />
Introduction xiii<br />
1 Image acquisition. Physical aspects. Instruments 1<br />
Introduction 1<br />
1.1 Mathematical model of the geometry of the aerial<br />
image 2<br />
YVES EGELS<br />
1.2 Radiometric effects of the atmosphere and the optics 16<br />
MICHEL KASSER<br />
1.3 Colorimetry 25<br />
YANNICK BERGIA, MICHEL KASSER<br />
1.4 Geometry of aerial and spatial pictures 34<br />
MICHEL KASSER<br />
1.5 <strong>Digital</strong> image acquisition with airborne CCD cameras 39<br />
MICHEL KASSER<br />
1.6 Radar images in photogrammetry 47<br />
LAURENT POLIDORI<br />
1.7 Use of airborne laser ranging systems for the<br />
determination of DSM 53<br />
MICHEL KASSER<br />
1.8 Use of scanners for the digitization of aerial pictures 58<br />
MICHEL KASSER<br />
1.9 Relations between radiometric and geometric precision<br />
in digital imagery 63<br />
CHRISTIAN THOM<br />
Architectural photogrammetry v
vi Contents<br />
2 Techniques for plotting digital images 78<br />
Introduction 78<br />
2.1 Image improvements 79<br />
ALAIN DUPÉRET<br />
2.2 Compression of digital images 100<br />
GILLES MOURY<br />
2.3 Use of GPS in photogrammetry 115<br />
THIERRY DUQUESNOY, YVES EGELS, MICHEL KASSER<br />
2.4 Automatization of aerotriangulation 124<br />
FRANCK JUNG, FRANK FUCHS, DIDIER BOLDO<br />
2.5 <strong>Digital</strong> photogrammetric workstations 145<br />
RAPHAËLE HENO, YVES EGELS<br />
3 Generation of digital terrain and surface models 159<br />
Introduction 159<br />
3.1 Overview of digital surface models 159<br />
NICOLAS PAPARODITIS, LAURENT POLIDORI<br />
3.2 DSM quality: internal and external validation 164<br />
LAURENT POLIDORI<br />
3.3 3D data acquisition from visible images 168<br />
NICOLAS PAPARODITIS, OLIVIER DISSARD<br />
3.4 From the digital surface model (DSM) to the digital<br />
terrain model (DTM) 221<br />
OLIVIER JAMET<br />
3.5 DSM reconstruction 225<br />
GRÉGOIRE MAILLET, PATRICK JULIEN,<br />
NICOLAS PAPARODITIS<br />
3.6 Extraction of characteristic lines of the relief 253<br />
ALAIN DUPÉRET, OLIVIER JAMET<br />
3.7 From the aerial image to orthophotography: different<br />
levels of rectification 282<br />
MICHEL KASSER, LAURENT POLIDORI<br />
3.8 Production of digital orthophotographies 288<br />
DIDIER BOLDO<br />
3.9 Problems relating to orthophotography production<br />
DIDIER MOISSET 292
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
4 Metrologic applications of digital photogrammetry 300<br />
Introduction 300<br />
4.1 Architectural photogrammetry 300<br />
PIERRE GRUSSENMEYER, KLAUS HANKE,<br />
ANDRÉ STREILEIN<br />
4.2 Photogrammetric metrology 340<br />
MICHEL KASSER<br />
Contents vii<br />
Index 349
viii Pierre Grussenmeyer, Klaus Hanke, André Streilein<br />
Plates<br />
There is a colour plate section between pages 192 and 193<br />
Figure 3.3.20 Left and right images of a 1m ground pixel size satellite<br />
across track stereopair; DSM result with a classical crosscorrelation<br />
technique; DSM using adaptive shape windows<br />
Figure 3.3.24 An example of the management of hidden parts<br />
Figure 3.3.27 Correlation matrix corresponding to the matching of the<br />
two epipolar lines appearing in yellow in the image extracts<br />
Figure 3.3.28 Results of the global matching strategy on a complete<br />
stereopair overlap in a dense urban area<br />
Figure 3.4.1 Images from the digital camera of IGN-F (1997) on Le<br />
Mans (France)<br />
Figure 3.8.3 Scanned images before balancing<br />
Figure 3.8.4 Scanned images after balancing<br />
Figure 3.9.2 Example from a real production problem: the radiometric<br />
balancing on digitized pictures<br />
Figure 3.9.4 Examples from the Ariège, France: problems posed by cliffs<br />
Figure 3.9.5 Generation of stretched pixels<br />
Figure 3.9.6 Example of radiometric difficulties (Ariège, France)<br />
Figure 3.9.7 Example of typical problems on water surfaces<br />
Figure 3.9.8 Two examples from the Isère department, France<br />
Figure 3.9.9 Two examples, using the digital aerial camera of IGN-F
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Contributors<br />
Architectural photogrammetry ix<br />
Yannick Bergia, born in 1977, graduated as an informatics engineer from<br />
ISIMA (Institut Supérieur d’Informatique, de Modélisation et de leurs<br />
Applications, Clermont-Ferrand, France) in 1999. He performed his<br />
diploma work in MATIS in 1999, and is a specialist in image processing.<br />
y.bergia@infonie.fr<br />
Didier Boldo, born in 1973, graduated from the École Polytechnique de<br />
Paris and from ENSG. He is now pursuing a Ph.D. in pattern recognition<br />
and image analysis at MATIS. His main research interests are<br />
photogrammetry, 3D reconstruction and radiometric modelling and calibration.<br />
didier.boldo@ign.fr<br />
Olivier Dissard, born in 1967, graduated from the École Polytechnique de<br />
Paris and from ENSG. Previously in charge of research concerning 3D<br />
urban topics at MATIS laboratory, he is now in charge of digital<br />
orthophotography at IGN-F. His studies at MATIS have concerned<br />
urban DEM and DSM, focusing on raised structures and classification<br />
in buildings and vegetation, and true orthophotographies.<br />
olivier.dissard@ign.fr<br />
Alain Dupéret graduated as ingénieur géographe from ENSG. He started<br />
working as a surveyor in IGN-F in 1979 and undertook topographic<br />
missions in France and Africa before specialising in image processing<br />
and DTM production. After managing technical projects and giving<br />
lectures in photogrammetry and image processing in ENSG, he became<br />
Head of Studies Management at ENSG.<br />
duperet@ensg.ign.fr<br />
Thierry Duquesnoy (born in 1963) graduated from ENSG in 1989. He<br />
gained his Ph.D. in earth sciences in 1997 at the University of Paris.<br />
He has worked since 1989 with the LOEMI, where he is a specialist in<br />
GPS trajectography for photogrammetry.<br />
thierry.duquesnoy@ign.fr
x Contributors<br />
Yves Egels, born in 1947, graduated from the École Polytechnique de Paris<br />
and from ENSG, and is senior photogrammetrist at IGN-F, where he<br />
pioneered analytical plotter software in the 1980s and digital workstations<br />
on PCs in the 1990s. He is now Head of the Photogrammetric<br />
Department at ENSG and gives lectures in photogrammetry at various<br />
French technical universities.<br />
egels@ensg.ign.fr<br />
Frank Fuchs, born in 1971, graduated from the École Polytechnique de<br />
Paris and from ENSG. Since 1996 he has been working at MATIS on<br />
his Ph.D. concerning the automatic reconstruction of buildings in aerial<br />
imagery through a structural approach.<br />
frank.fuchs@ign.fr<br />
Pierre Grussenmeyer, born in 1961, graduated from ENSAIS, and gained a<br />
Ph.D. in photogrammetry in 1994 at the University of Strasbourg in<br />
collaboration with IGN-F. He has been on the academic staff of the<br />
Department of Surveying at ENSAIS, where he teaches photogrammetry,<br />
since 1989. Since 1996 he has been Professor and the Head of the<br />
<strong>Photogrammetry</strong> and Geomatics Group at ENSAIS-LERGEC.<br />
pierre.grussenmeyer@ensais.u-strasbg.fr<br />
Klaus Hanke, born 1954, studied geodesy and photogrammetry at the<br />
University of Innsbruck and the Technical University of Graz. In 1984<br />
he became Dr. techn., and since 1994 he has been a Professor teaching<br />
<strong>Photogrammetry</strong> and Architectural <strong>Photogrammetry</strong> at the University<br />
of Innsbruck.<br />
Raphaële Heno, born 1970, graduated in 1993 from ENSG and from ITC<br />
(Enschede – The Netherlands). She spent four years developing digital<br />
photogrammetric tools to improve IGN-France topographic database<br />
plotting. Since 1998 she has been senior consultant for the IGN-France<br />
advisory department.<br />
raphaele_heno@hotmail.com<br />
Olivier Jamet, born in 1963, graduated from the École Polytechnique de<br />
Paris and from ENSG. He gained a Ph.D. in signal and image processing<br />
at the École Nationale Supérieure des Télécommunications de Paris<br />
(1998). Formerly Head of the MATIS laboratory, he is currently in<br />
charge of research in physical geodesy at LAREG, ENSG.<br />
olivier.jamet@ign.fr<br />
Patrick Julien graduated from the University of Paris (in mathematics) and<br />
from ENSG, and joined IGN-F in 1975. He has developed various softwares<br />
for orthophotography, digital terrain models and digital image<br />
matching. He is presently a researcher at MATIS and gives lectures at<br />
ENSG in geographic information sciences.<br />
patrick.julien@ign.fr
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Contributors xi<br />
Franck Jung, born in 1970, graduated from the École Polytechnique de<br />
Paris and from ENSG, and is currently a researcher at the MATIS laboratory.<br />
He is preparing a Ph.D. dissertation on automatic change<br />
detection using aerial stereo pairs at two different dates. His main<br />
research interests are automatic digital photogrammetry, change detection<br />
and machine learning.<br />
franck.jung@ign.fr<br />
Michel Kasser, born in 1953, graduated from the École Polytechnique de<br />
Paris and from ENSG. He created the LOEMI in 1984 and was from<br />
1991 to 1999 head of the ESGT. As a specialist in instrumentation, in<br />
geodesy and in space imagery, he is University Professor and Head of<br />
the Geodetic Department and of LAREG at IGN-F.<br />
michel.kasser@ign.fr<br />
Grégoire Maillet, born in 1973, graduated from ENSG in 1997. He is now<br />
a research engineer at MATIS, where he works on dynamic programming<br />
and automatic correlation for image matching.<br />
gregoire.maillet@ign.fr<br />
Didier Moisset, born in 1958, obtained the ENSG engineering degree in<br />
1983. He was in charge of the aerotriangulation team at IGN in 1989,<br />
and was Professor in photogrammetry and Head of the Photogrammetric<br />
Department at the ENSG during 1993–8. He was a project manager in<br />
1998 and developed new automatic orthophoto software for IGN. He<br />
was recently appointed as a consultant in photogrammetry at the MATIS<br />
laboratory.<br />
didier.moisset@ign.fr<br />
Gilles Moury, born in 1960, graduated from the École Polytechnique de<br />
Paris and from ENSAE (École Nationale Supérieure de l’Aéronautique<br />
et de l’Espace). He has been working in the field of satellite on-board<br />
data processing since 1985, within CNES. In particular, he was responsible<br />
for the development of image compression algorithms for various<br />
space missions (Spot 5, Clementine, etc.). He is now Head of the onboard<br />
data processing section and lectures in data compression at various<br />
French technical universities.<br />
gilles.moury@cnes.fr<br />
Nicolas Paparoditis, born in 1969, obtained a Ph.D. in computer vision<br />
from the University of Nice-Sophia Antipolis. He worked for five years<br />
for the Aérospatiale Company on the specification of HRV digital<br />
mapping satellite instruments (SPOT satellites). He is a specialist in<br />
image processing, in computer vision, and in digital aerial and satellite<br />
imagery, and is currently a Professor at ESGT where he leads the research<br />
activities in digital photogrammetry. He is also a research member of<br />
MATIS at IGN-F.<br />
nicolas.paparoditis@ign.fr
xii Contributors<br />
Laurent Polidori, born in 1965, graduated from ENSG in 1987. He gained<br />
a Ph.D. in 1991 on ‘DTM quality assessment for earth sciences’ at Paris<br />
University. A researcher at the Aérospatiale Company (Cannes, France)<br />
on satellite imagery, he is the author of Cartographie Radar (Gordon<br />
and Breach, 1997). He is Head of the Remote Sensing Laboratory of<br />
IRD (Cayenne, Guyane) and Associate Professor at ESGT.<br />
polidori@caiena.cayenne.ird.fr<br />
André Streilein in 1990 obtained a Dip.Ing. at the Rheinische Friedrich-<br />
Wilhelms-Universität, Bonn. In 1998 he obtained a Dr. sc. tech. at the<br />
Swiss Federal Institute of Technology, Zurich. His dissertation was on<br />
‘<strong>Digital</strong>e Photogrammetrie und CAAD’. Since September 2000 he has<br />
been at the Swiss Federal Office of Topography (Bern), Section<br />
<strong>Photogrammetry</strong> and Remote Sensing. (Postal address: Bundesamt für<br />
Landestopographie, Seftigenstrasse 2, 3084 Wabern, Bern, Switzerland.)<br />
Christian Thom, born in 1959, graduated from the École Polytechnique<br />
de Paris and from the ENSG. With a Ph.D. (1986) in robotics and signal<br />
processing, he has specialized in instrumentation. He is Head of LOEMI<br />
where he pioneered and developed the IGN program of digital aerial<br />
cameras from 1989.<br />
christian.thom@ign.fr<br />
Institutions<br />
IGN-F is the Institut Géographique National, France, administration in<br />
charge of national geodesy, cartography and geographic databases. Its main<br />
installation is in Saint-Mandé, close to Paris. The MATIS is the laboratory<br />
for research in photogrammetry of IGN-F; the LOEMI is its laboratory<br />
for new instrument development; and the LAREG its laboratory for<br />
geodesy. The ENSG (École Nationale des Sciences Géographiques) is the<br />
technical university owned by IGN-F.<br />
(Postal address: IGN, 2 Avenue Pasteur, F-94 165 Saint-Mandé Cedex,<br />
France.)<br />
The CNES is the French Space Agency.<br />
(Postal address: CNES, 2 Place Maurice Quentin, F-75 039 Paris Cedex<br />
01, France.)<br />
The ENSAIS is a technical university in Strasbourg with a section of engineers<br />
in surveying.<br />
(Postal address: ENSAIS, 24 Boulevard de la Victoire, F-67 000, Strasbourg,<br />
France.)<br />
The ESGT is the École Supérieure des Géomètres et Topographes<br />
(Le Mans), the technical university with the most important section of<br />
engineers in surveying in France.<br />
(Postal address: ESGT, 1 Boulevard Pythagore, F-72 000, Le Mans, France.)
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Introduction<br />
Architectural photogrammetry xiii<br />
In 1998, when Taylor & Francis pushed me into the adventure of writing<br />
this book with my colleague Yves Egels, most specialists of photogrammetry<br />
around me were aware of the lack of communication between their<br />
technical community and other specialists. Indeed, this situation was not<br />
new, as 150 years after its invention by Laussédat photogrammetry was<br />
really reserved for photogrammetrists (considering for example the<br />
extremely high costs of photogrammetric plotters, but also the necessary<br />
visual ability of the required technicians, a skill that required a long and<br />
costly training, etc.). <strong>Photogrammetry</strong> was thus the speciality of a very<br />
small club, mostly European, and the related industrial activity was heavily<br />
based on the national cartographic institutions.<br />
As we saw it in the early 1990s, the situation has changed very rapidly<br />
with the improvement of the computational power available on cheap<br />
machines. First, one could see software developments trying to avoid the<br />
mechanical part of the photogrammetric plotters, but more or less based<br />
on the same algorithms as analytical plotters. These machines, still devoted<br />
to professional practitioners, appeared a little cheaper, but with a comparatively<br />
poor visual quality. They were often considered as good workstations<br />
for orthophotographic productions with, as a side quality, an additional<br />
production capacity of photogrammetric plotting in case of emergency<br />
work. The main residual difficulty was linked to the huge amount of bytes,<br />
and also to the digitization of the traditional argentic aerial images, as the<br />
necessary scanners were extremely expensive. The high cost of the very precise<br />
mechanical parts had already been reduced in the 1980s from the old<br />
opto-mechanical plotters to the analytical plotters, as the mechanics has<br />
been downsized from 3D to 2D. But as a final revolution they had just<br />
shifted from the analytical plotter to the scanner, although a scanner could<br />
be used for many plotting workstations, and thus the total production cost<br />
was already lower. But this necessity to use a high-cost scanner was the<br />
main limitation of any new development of photogrammetric activity.<br />
During this period, the industry developed for the consumer public very<br />
interesting products devoted to following the explosion of the microinformatics<br />
evolution and providing a set of extremely cheap tools: video
xiv Introduction<br />
cards allowing a fast and nearly real-time image management, stereoscopic<br />
glasses for video games, A4 scanners, digital colour cameras, etc. These<br />
developments were quickly taken into account by photogrammetrists, so<br />
that the <strong>Digital</strong> Photogrammetric Workstations (DPW) became cheaper<br />
and cheaper, most of the cost today being the software cost. This has<br />
completely changed the situation of photogrammetry: it is now possible<br />
to own a DPW and to use it only occasionally, as the acquisition costs<br />
are low. And if the user lacks a good stereoscopic vision capability, the<br />
automatic matching will help him to put the floating point in good contact<br />
with the surveyed zone. Of course, only trained people may quickly survey<br />
a large amount of data in a cost-effective way, but a good possibility to<br />
work is opened to non-specialists.<br />
If the situation changed a lot for non-photogrammetrists, in fact it<br />
changed a lot for the photogrammetrists themselves, too: the main revolution<br />
we have met in the 1990s is the availability of aerial digital cameras,<br />
providing images whose quality is far superior to the classic argentic ones,<br />
at least in terms of radiometric fidelity, noise level, and geometric perfection.<br />
Thus the last obstacle for a wide spreading of digital photogrammetry<br />
is disappearing. We are on the verge of a massive transfer of such techniques<br />
toward purely digital processes, and the improvement in quality of<br />
the products, hopefully, should be very significant in the next few years.<br />
Also, the use of digital images allows us to automate more and more tasks,<br />
even if a large part of the automation is still at a research level (most automatic<br />
cue extractions).<br />
We have then a completely new situation, where (for a basic activity at<br />
least) the equipment is no longer a limit for a wide spread of photogrammetric<br />
methodology. But if the use is now spreading rapidly, the theoretical<br />
aspects are not sufficiently known today, which could result in disappointing<br />
results for newcomers.<br />
Thus this book has been oriented towards engineers and technicians,<br />
including technicians who are basically external to the field of photogrammetry<br />
and who would like to understand at least part of this field, for<br />
example to use it from time to time, without the help of professionals of<br />
photogrammetry. The goal here is to get this specialty much closer to the<br />
end user, and to help him to be autonomous in most circumstances, a little<br />
like word processors allowed us to type what we wanted when we wanted<br />
it, twenty years ago. But this end user must already have a scientific and<br />
technical culture, as the theoretical concepts would not be accessible<br />
without such a background. So I have supposed in this publication that<br />
the reader had such an initial understanding. Of course, professors and<br />
students, young ones as well as former specialists going through Continuous<br />
Professional Development, will also find great benefit from this up-to-date<br />
material.<br />
A final very important point: the goal here being to be extremely close<br />
to the practical applications and to the production of data, we have looked
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Introduction xv<br />
for co-authors who are, each of them, excellent specialists in given aspects<br />
of digital photogrammetry, and most of them are practitioners or directly<br />
working for them. This is a guarantee that the considerations discussed<br />
here are really close to the application world. But the unavoidable drawback<br />
is that we have here 21 different authors for a single book, and that<br />
despite our efforts it is obvious that the homogeneity is not perfect. Some<br />
aspects are presented twice, sometimes with significant differences of point<br />
of view: this will remind each of us that the reality is complex, even in<br />
technical matters, and that controversy is the essence of our societies. In<br />
any case we have included an abundant bibliography that may be useful<br />
for further reading. Of course, as many authors are from France, most of<br />
the illustrations come from studies performed in this country, especially<br />
as IGN-F has been one of the few pioneers of the digital CCD cameras.<br />
But I think that this material is more or less equivalent to what everybody<br />
may find in his own country, especially now that large companies commercially<br />
propose digital cameras.<br />
Michel Kasser
xvi Pierre Grussenmeyer, Klaus Hanke, André Streilein
1111<br />
2<br />
3<br />
4<br />
5111<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
44111<br />
1 Image acquisition. Physical<br />
aspects. Instruments<br />
INTRODUCTION<br />
We may define photogrammetry as ‘any measuring technique allowing the<br />
modelling of a 3D space using 2D images’. Of course, this is perfectly suitable<br />
for the case where one uses photographic pictures, but this is still the<br />
case when any other type of 2D acquisition device is used, for example a<br />
radar, or a scanning device: the photogrammetric process is basically independent<br />
of the image type. For that reason, we will present in this chapter<br />
the general physical aspects of the data acquisition in digital photogrammetry.<br />
It starts with §1.1, a presentation of the geometric aspects of<br />
photogrammetry, which is a classic problem but whose presentation is<br />
slightly different from books dealing with traditional photogrammetry. In<br />
§1.1 there are also some considerations about atmospheric refraction, the<br />
distortion due to the optics, and a few rules of thumb in planning aerial<br />
missions where digital images are used. In §1.2, we present some radiometric<br />
impacts of the atmosphere and of the optics on the image, which<br />
helps to understand the very physical nature of the aerial images. And in<br />
§1.3, these few lines about colorimetry are necessary to understand why<br />
the digital tools for managing black and white panchromatic images need<br />
not be the same as the tools for managing colour images. Then we will<br />
present the instruments used to obtain the images, digital of course. In<br />
§1.4 we will consider the geometric constraints on optical image acquisitions<br />
(aerial, and satellite as well). From §1.5 to §1.8, the main digital<br />
acquisition devices will be presented, CCD cameras (§1.5), radars (§1.6),<br />
airborne laser scanners (§1.7), and photographic image scanners (§1.8).<br />
And we will close this chapter with an analysis of the key problem of the<br />
size of the pixel, and the signal-to-noise ratio in the image.
2 Yves Egels<br />
1.1 MATHEMATICAL MODEL OF THE GEOMETRY OF<br />
THE AERIAL IMAGE<br />
Yves Egels<br />
The geometry of images is the basis of all photogrammetric processes,<br />
analogical, analytic or digital. Of course, the detailed geometry of an image<br />
essentially depends on the physical features of the sensor used. In analytic<br />
photogrammetry (the only difference with digital photogrammetry is the<br />
digital nature of the image) the use of this technology of image acquisition<br />
requires therefore a complete survey of its geometric features, and of<br />
the mathematical tools expressing the relation between the coordinates on<br />
the image and those of ground points (and, if the software is correct, this<br />
will be the only thing to do).<br />
We shall explain here, just as an example, the case of the traditional<br />
conical photography (the image being silver halide or digital), that is still<br />
today the most used and that will be able to be a guide for the analysis<br />
of a different sensor. The leading principle, here as well as in any other<br />
geometrical situation, is to approach the quite complex real geometry<br />
by a simple mathematical formula (here the perspective) and to consider<br />
the differences between the physical reality and this formula as corrections<br />
(‘corrections of systematic errors’) presented independently. The imaging<br />
devices whose geometry is different are presented later in this chapter<br />
(§§1.4 to 1.7).<br />
1.1.1 Mathematical perspective of a point of the space<br />
The perspective of a point M of the space, of centre S, on the plane P is<br />
the intersection m of the line SM with the plane P. Coordinates of M will<br />
be expressed in the reference system (O, X, Y, Z), and the ones of m in<br />
the reference system of the image (c, x, y, z). The rotation matrix R corresponds<br />
to the change of system (O, X, Y, Z) → (c, x, y, z). One will note<br />
F the coordinates of the summit S in the reference of the image.<br />
M XM YM ZM S XS YS ZS F xc yc m image of M ⇔ Fm → .SM → .<br />
p m x<br />
y<br />
If one calls K the vector unit orthogonal to the plane of the image one<br />
can write:<br />
0<br />
m ∈ picture ⇔ Kt K<br />
m 0 that implies ’<br />
tF K t R(M S)
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
X<br />
m<br />
Z<br />
O<br />
from where the basic equation of the photogrammetry, so-called collinearity<br />
equation:<br />
m F . (1.1)<br />
This equation can be developed and reordered according to coordinates<br />
of M, under the so-called ‘11 parameters equation’, which appears simpler<br />
to use, but should be handled with great care:<br />
KtFR(M S)<br />
K t R(M S)<br />
x a 1X M b 1Y M c 1Z M d 1<br />
a 3 X M b 3 Y M c 3 Z M 1<br />
K<br />
z<br />
c<br />
F<br />
S<br />
Y<br />
y<br />
x<br />
The geometry of the aerial image 3<br />
Figure 1.1.1 Systems of coordinates for the photogrammetry.<br />
R<br />
M<br />
y a 2X M b 2Y M c 2Z M d 2<br />
a 3 X M b 3 Y M c 3 Z M 1<br />
. (1.2)
4 Yves Egels<br />
1.1.2 Some mathematical tools<br />
1.1.2.1 Representation of the rotation matrixes, exponential and<br />
axiator<br />
In computations of analytic photogrammetry, the rotation matrixes appear<br />
quite often. On a vector space of n dimensions, they depend on n(n 1)/2<br />
parameters, but do not form a vector sub-space. There is therefore no<br />
linear formula between a rotation matrix and its ‘components’. It is,<br />
however, vital to be able to express such a matrix judiciously according<br />
to the chosen parameters. In the case of R3 , it is usual to analyse the rotation<br />
as the composition of three rotations around the three axes of<br />
coordinates (angles w, and , familiar to photogrammetrists). But the<br />
resulting formulas are laborious enough and lack symmetry, which obliges<br />
us to perform unnecessary developments of formulas. One will be able to<br />
define a rotation of the following manner physically: rotation of an angle<br />
around a unitary vector<br />
→ a<br />
b<br />
c , with (√a2 b2 c2 = 1)<br />
We will call here the axiator of a vector the matrix equivalent to a vectorial<br />
product:<br />
→ ∧ V → ~ ·V<br />
one will check that<br />
~ 0<br />
c<br />
b<br />
. (1.3)<br />
It is easy to control that the previous rotation can be written: R e ~ .<br />
Indeed,<br />
, (1.4)<br />
therefore, since ~ · → ∧ → 0 (a well-known property of the vectorial<br />
product) e ~ → →<br />
·. is the rotation axis, only invariant. One will<br />
check that ~ e<br />
~ ~ 3 4 , 2, ~ ~ t etc.; on the other hand, ~ ~ <br />
. . . ~ n n <br />
. . .<br />
n!<br />
(e ~ )<br />
c<br />
0<br />
a<br />
b<br />
a<br />
0<br />
~ t . . . ~ t n<br />
n!<br />
n<br />
. . .
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
<br />
(1.5)<br />
.<br />
Therefore this matrix is certainly an orthogonal matrix, which is a rotation<br />
matrix as det R 1 (as ~ e<br />
~ t ). One will also verify that the rotation<br />
angle is equal to , for example by computing the scalar product of a<br />
normal vector to with its transform.<br />
One sees therefore, while regrouping the even terms and the odd terms<br />
of the procedure:<br />
~ (e ~ 1 )<br />
R ~ ~ 3<br />
~ . . . (1) ~ n n<br />
R I . (1.6)<br />
~ sin ~ 2 (1 cos ) (Euler formula)<br />
This formula will permit a very comfortable calculation of rotations. Indeed<br />
one can choose, as parameters of the rotation, the three values a, b and<br />
c. One can then write:<br />
→ →<br />
Θ . (1.7)<br />
a<br />
c and thus ||Θ→ ||, → Θ→<br />
b<br />
If one expresses sin and cos using tan /2 with the help of the classic<br />
trigonometric formulas, it becomes:<br />
R (I . (1.8)<br />
~ tan (/2)) 1 (I ~ tan (/2)) (Thomson formula)<br />
1.1.2.2 Differential of a rotation matrix<br />
Expressed under this exponential shape, it is possible to differentiate the<br />
rotation matrixes. One may determine:<br />
dR ≈ e Θ~ dΘ ~ R dΘ ~<br />
3! . . . ~ 2 2 <br />
2! ~ 2 4 <br />
4!<br />
The geometry of the aerial image 5<br />
n!<br />
. . .<br />
. . .<br />
. (1.9)<br />
In this expression, the differential makes the parameters of the rotation<br />
appear under matrix shape, which is not very practical. Applied to a vector,<br />
it can be written (using the anticommutativity of the vectorial product):<br />
dR·A R·dΘ . (1.10)<br />
~ ·A R·A ~ ·dΘ
6 Yves Egels<br />
1.1.2.3 Differential relation binding m to the parameters of the<br />
perspective<br />
The previous equation is not linear, and to solve systems corresponding<br />
to analytical photogrammetry (orientation of images, aerotriangulation),<br />
it is necessary to linearize them. Variables that will be taken into account<br />
are F, M, S and R.<br />
If one puts A M S and U RA<br />
(1.11)<br />
dU R(dM dS) dRA R(dM dS A ~ dm dF <br />
d). (1.12)<br />
KtdF U<br />
KtU KtF KtU KtFUKt (Kt 2 U) dU<br />
On the other hand, K t F being a scalar, K t FU UK t F and K t dFU UK t dF.<br />
One then gets:<br />
dm . (1.13)<br />
If one writes<br />
KtU UKt (KtU) 2 [KtU dF KtFR(dM dS A ~ dΘ)]<br />
p KtF, U u1 u2 u3 and V u3 0 0<br />
u3 one gets after manipulation:<br />
dx V<br />
<br />
dy u3 dF p<br />
2 u3 . (1.14)<br />
Under this matrix shape, the computer implementation of this equation<br />
requires only a few code lines.<br />
1.1.3 True geometry of images<br />
u 1<br />
u 2 ,<br />
p<br />
VR(dS dM) 2 u3 VR A~ dΘ<br />
In physical reality, the geometry of images only reproduces in an approximate<br />
way the mathematical perspective whose formulas have just been<br />
established. If one follows the light ray from its starting point on the<br />
ground until the formation of the image on the sensor, one will find a<br />
certain number of differences between the perspective model and the reality.<br />
It is difficult to be comprehensive in this area, and one will mention here<br />
only the main reasons for distortion whose consequences are appreciable<br />
in the photogrammetric process.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
1.1.3.1 Correction of Earth curvature<br />
The first reason for which photogrammetrists use the so-called correction<br />
of Earth curvature in photogrammetry is that the space in which cartographers<br />
operate is not the Euclidian space in which the collinearity<br />
equation is set up. It is indeed in a cartographic system, in which the<br />
planimetry is measured in a plane representation (often conform) of the<br />
ellipsoid, and the altimetry in a system of altitudes.<br />
The best solution for dealing with this is to make rigorous transformations<br />
from cartographic to tridimensional frames (it is usually more<br />
convenient to choose a local tangent tridimensional reference, if it is necessary<br />
just to write relations of planimetric or altimetric conversion in a<br />
simple way). This solution, if better in theory, is, however, a little difficult<br />
to put into an industrial operation, because it requires having a<br />
database of geodesic systems, of projection algorithms (sometimes bizarre)<br />
used in the client countries, and arranging a model of the zero level surface<br />
(close to the geoid) of the altimetric system concerned.<br />
An approximate solution to this problem can nevertheless be found, that<br />
is – experience proves it – of comparable precision with the precision of<br />
this measurement technique. It relies on the fact that all conform projections<br />
are locally equivalent, with only a changing scale factor. One can<br />
attempt to replace the real projection, in which is expressed the coordinates<br />
of the ground, by a unique conform projection therefore (and why<br />
not try for simplicity!), for example a stereographic oblique projection of<br />
the mean curvature sphere on the tangent plane to the centre of the work.<br />
The mathematical formulation then becomes as shown in Figure 1.1.2<br />
(the figure is made in the diametrical plane of the sphere containing the<br />
centre of the work A and the point I to transform).<br />
I is the point to transform (cartographic), J the result (tridimensional),<br />
I′ the projection of I on the plane and J′ the projection of J on the terrestrial<br />
sphere of R radius. I′ and J′ are the inverse of each other in an<br />
inversion of B pole and coefficient 4R2 .<br />
Let us put:<br />
I x<br />
Rh I′ x<br />
R J′ x′<br />
y′ J (1 ) x′<br />
y′<br />
The calculation of the inversion gives:<br />
||BI′|| · ||BJ′|| 4R 2<br />
BJ′ BI′ <br />
J′ B B′ x(12 )<br />
4R2<br />
⇒ 2 <br />
BI′<br />
The geometry of the aerial image 7<br />
x<br />
1<br />
2 ≈ 1 2<br />
1 <br />
R(12 2 ) J (12 )(1)x<br />
(12 2 )(1)R<br />
2R<br />
h<br />
.<br />
R<br />
(1.15)
8 Yves Egels<br />
O<br />
JI (2 )x<br />
2 2 R<br />
. (1.16)<br />
One can consider that J I corresponds to a correction to bring us to the<br />
coordinates of I:<br />
I curvature xh/Rx3 /4R 2<br />
x 2 /2R<br />
B<br />
A<br />
R<br />
Figure 1.1.2 Figure made in the diametrical plane of the sphere containing the<br />
centre of the work A and the point I to transform.<br />
.<br />
J′<br />
(1.17)<br />
In the current conditions, the term of altimetric correction is by far the<br />
most meaningful. But for aerotriangulations of large extent, planimetric<br />
corrections can become important, and must be considered. If the aerotriangulation<br />
relies on measures of an airborne GPS (global positioning<br />
system), it is necessary not to forget also to do this correction on coordinates<br />
of tops, that also have to be corrected for the geoid–ellipsoid<br />
discrepancy.<br />
1.1.3.2 Atmospheric refraction<br />
Before penetrating the photographic camera, the luminous rays cross the<br />
atmosphere, whose density, and therefore refractive index, decreases with<br />
the altitude (see Table 1.1.3).<br />
h<br />
J<br />
I'<br />
I<br />
h<br />
x
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
This variation of the index provokes a curvature of the luminous rays<br />
(oriented downwards) whose amplitude depends on atmospheric conditions<br />
(pressure and temperature at the time of the image acquisition), that<br />
are not uniform in the field of the camera, and are generally unknown.<br />
Nevertheless, this deviation is small with regard to the precision of photogrammetric<br />
measurements (with the exception of very small scales). It will<br />
thus be acceptable to correct the refraction using a standard reference<br />
atmosphere: in order to evaluate the influence of the refraction, it will be<br />
necessary to use a model providing the refraction index at various altitudes.<br />
Numerous formulas exist allowing an evaluation of the variation of the<br />
air density according to altitude. No one is certainly better that any other,<br />
because of the strong turbulence of low atmospheric layers. In the tropospheric<br />
domain (these formulas would not be valid if the sensor is situated<br />
in the stratosphere), one will be able to use, for example, the formula<br />
published by the OACI (Organisation de l’Aviation Civile Internationale):<br />
where<br />
n n 0 (1 ah bh 2 ), (1.18)<br />
a 2.560.10 8 , b 7.5.10 13 , n 0 1.000278 h in metres,<br />
or this one, whose integration is simpler:<br />
where<br />
n 1 ae bh , (1.19)<br />
a 278.10 6 , b 105.10 6 .<br />
The geometry of the aerial image 9<br />
Table 1.1.3 Atmospheric co-index of refraction–variations with the altitude<br />
H (km) 1 0 1 3 5 7 9 11<br />
N (n1)·10 6 306 278 252 206 167 134 106 83<br />
The rays appear to come then from a point situated in the extension of<br />
the tangent to the ray at the entrance into the optics, thus introducing a<br />
displacement of the image.<br />
The calculation of the refraction correction can be performed in the<br />
following way: the luminous ray coming from M seems to originate from<br />
the point M′ situated on one same vertical, so that MM′ y. One will<br />
modify the altitude of the object point M by the correction y by putting<br />
it in M′. This method of calculation has the advantage of not supposing
10 Yves Egels<br />
M′<br />
M<br />
y<br />
that the axis of the acquired image is vertical, and thus works also in<br />
oblique terrestrial photogrammetry. (See Figure 1.1.4.)<br />
dy h hM dv ,<br />
sin v · cos v<br />
from where<br />
S<br />
y MM′ M<br />
. (1.20)<br />
The formula of Descartes gives us the variation of the refraction angle:<br />
n · sin v cte,<br />
or dn sin v n cos v dv 0,<br />
or again<br />
d<br />
S<br />
h<br />
v<br />
Figure 1.1.4 Optical ray in the atmosphere (with a strong exaggeration of the<br />
curvature).<br />
h h M<br />
sin v · cos v<br />
dv . (1.21)<br />
In the case of photography of a zone of small extension, one will suppose<br />
the Earth to be flat, the vertical (and the refractive index gradients) are<br />
dn<br />
tan v<br />
n<br />
x
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
therefore all parallel to each other (this approximation is not valid in<br />
spatial imagery). The radius being very large, one will be able to assimilate<br />
the bow and the chord, the v angle and the refractive index can be<br />
considered constant for the integration.<br />
MM′ <br />
S<br />
M<br />
S 1<br />
(h h .<br />
M) (1.22)<br />
dn<br />
dh dh<br />
cos2 v M<br />
With this second formula for refraction, the calculation for the complete<br />
rays gives:<br />
dn<br />
dh ab ebh , dn<br />
dh dh a ebh n 1<br />
and<br />
h .<br />
(1.23)<br />
If one puts D ||SM|| and H hS hM one gets the following refraction<br />
correction:<br />
dn<br />
dh dh 1<br />
h (n 1) b<br />
y D2<br />
H2 nS nM H(nS 1) b<br />
.<br />
(1.24)<br />
For aerial pictures at the 1/50,000th scale with a focal length of 152 mm,<br />
this correction is of about 0.20 m to the centre, and reaches 0.50 m at<br />
the edge of the image. It is appreciably less important at larger scales for<br />
the same focal length.<br />
This formula also permits either the correction of refraction with oblique<br />
or horizontal optical axes lines (terrestrial photogrammetry), which are<br />
very obviously not acceptable domains in very commonly used formulas,<br />
bringing a radial correction to the image coordinates. One will be able to<br />
verify that when h M → h S, the limit of y is:<br />
y H n S 1 b<br />
2 D2 ,<br />
(h hM) sin v<br />
sin v · cos v cos v dn<br />
n<br />
The geometry of the aerial image 11<br />
S<br />
M<br />
(h hM) cos 2 v dn<br />
n<br />
the classical refraction formula for a horizontal optical axis.
12 Yves Egels<br />
1.1.3.3 At the entrance in the plane<br />
In a pressurized plane two phenomena may occur. First, the light ray<br />
crosses the glass plate placed before the objective. This glass plate, therefore<br />
of weak thermal conductivity, is subject to significant temperature<br />
differences and bends under the differential dilation effect (in the opposite<br />
direction to the bending due to the pressure difference, that induces<br />
a much smaller effect). It takes a more or less spherical shape, and operates<br />
then like a lens that introduces an optic distortion. Unfortunately, if<br />
the phenomenon can be experimentally tested (and the calculation shows<br />
that its influence is not negligible), it is practically impossible to quantify<br />
it, the real shape of the glass plate porthole depends on the history of the<br />
flight (temperatures, pressures inside and outside). There are techniques<br />
of calculation that will be applied to §1.1.3.6, however, (a posteriori<br />
unknowns) permitting one to model and to eliminate the influence of it<br />
in the aerotriangulation.<br />
When the plane is pressurized the luminous ray is refracted a second<br />
time at the entrance to the cabin, where the pressure is higher than that<br />
outside at the altitude of flight (the difference in pressure is a function<br />
of the airplane used). One can model this refraction in the same way as<br />
for the atmospheric refraction (the sign is opposed to the previous one,<br />
because the radius arrives in a superior index medium).<br />
While noting nc the refractive index inside the cabin, and while supposing<br />
that the porthole is horizontal,<br />
dv (n c n s ) tan v (n c n s ) R<br />
H<br />
dy D2<br />
H 2 (n c n s)H.<br />
(1.25)<br />
(1.26)<br />
While combining this equation with that of atmospheric refraction, one<br />
sees that it is sufficient to replace the altitude of the top by the altitude<br />
of the cabin in the second part:<br />
y D2<br />
H2 nS nM H(nc 1) b<br />
.<br />
(1.27)<br />
Thus we have to compute the cabin index of refraction. Often in pressurized<br />
planes, the pressure is limited to a maximal value of P max (for<br />
example 6 psi 41,370 Pa for the Beechcraft King Air of IGN-F) between<br />
inside and outside. Under a given altitude limit (around 4,300 m in the<br />
previous example), the pressure is kept to the ground value. The refractive<br />
index n and co-index N are linked to the temperature and pressure<br />
by an approximate law:
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
N (n 1)106 790 , P in kPa and T in K.<br />
The temperature decreases with the altitude with an empirical variation T<br />
T0 6.5 × 103 P<br />
T<br />
Z. If one takes TC T0 298 K, the cabin refractive<br />
index will be:<br />
N C N S (1 2.22 × 10 5 Z S) 2,696 P.<br />
The geometry of the aerial image 13<br />
This index will be limited to its ground value (N 278 in this model),<br />
if the altitude is below the pressurization limit. One may note that, in<br />
current flight conditions (flight altitude 5,000 m, pressurization limit<br />
4,300 m), this total correction is nearly the opposite of the correction<br />
computed without cabin pressurization.<br />
1.1.3.4 Distortion of the optics<br />
In elementary optics (conditions of Gauss: thin lenses, rays with a low tilt<br />
angle with respect to the optical axis, low aperture optics), all rays passing<br />
through the centre of the objective are not deviated. This replicates the<br />
mathematical perspective. Unfortunately, the real optics used in photogrammetric<br />
image acquisitions do not fulfil these conditions.<br />
In the case of perfect optics (centred spherical dioptres) the entrance<br />
ray, the exit ray, and the optical axis are coplanar. But the incident and<br />
emergent rays are not parallel. This gap of parallelism constitutes the<br />
distortion of the objective. It is a permanent and steady characteristic of<br />
the optics, which can be measured prior to the image acquisitions, and<br />
corrected at the time of photogrammetric calculations.<br />
The measure of this distortion can be achieved on an optical bench,<br />
using collimators, or by observation of a dense network of targets through<br />
the objective. Manufacturers, to provide certification of the aerial camera<br />
that they produce, use these methods, which are quite complicated and<br />
costly. It is also possible to determine the distortion (as well as the centring<br />
and the principal distance) by taking images of a tridimensional polygon<br />
of points very well known in coordinates, and by calculation of the<br />
collinearity equations, to which one adds unknowns of distortion according<br />
to the chosen model.<br />
In the case of perfect optics, the distortion is very correctly modelled<br />
by a radial symmetrical correction polynomial around the main point of<br />
symmetry (intersection of the optic axis with the plane of the sensor).<br />
Besides, one can demonstrate easily that this polynomial only contains<br />
some terms of odd rank. The term of first degree can be chosen arbitrarily,<br />
because it is bound by choice of the main distance: one often takes it as<br />
a null (main distance to the centre of the image) or so that the maximal<br />
distortion in the field is the weakest possible.
14 Yves Egels<br />
If the optics cannot be considered as perfect (it is sometimes the case if<br />
one attempts to use cameras designed for the general public for photogrammetric<br />
operations), the distortion can lose its radial and symmetrical<br />
character, and requires a means of calibration and more complex modelling.<br />
Besides, in the case of zooms (that one should avoid . . .), this distortion<br />
does not remain constant in time because of variations of centrage brought<br />
about by the displacement of the pancratic vehicle (mobile part permitting<br />
the displacement of lenses). Any modelling then becomes very uncertain.<br />
1.1.3.5 Distortions of the sensor<br />
The equation of the perspective supposes that the sensor is a plane, and<br />
that one can define there a reference of fixed measure. Thus, it will be<br />
necessary to compare the photochemical sensors (photographic emulsion)<br />
and the electronic sensors (CCD matrixes). These last are practically indeformable,<br />
and the image that they acquire is definitely steady (in any case<br />
on the geometric plane). The following remarks will therefore apply to the<br />
photochemical sensors only.<br />
The planarity of the emulsion was formerly obtained by the use of photographic<br />
glass plates, but this solution is no longer used, except maybe in<br />
certain systems of terrestrial photogrammetry. It was, besides, not very<br />
satisfactory. Today one practically always uses vacuum depression of the<br />
film on the plate at the bottom of the camera. This method is very convenient,<br />
provided that there is no dust interposed between the film and<br />
the plate, which is unfortunately quite difficult to achieve. There follows<br />
a local distortion of the image, which is difficult to detect (except in radiometrically<br />
uniform zones, where there appears an effect of shade) and<br />
impossible to model.<br />
The dimensional stability of the image during the time, if it is improved<br />
with the use of so-called ‘stable’ supports, is quite insufficient for the use<br />
of photogrammetrists. Distortions ten times better than the necessary<br />
measurement precision are frequent. The only efficient counter-measure is<br />
the measure of the coordinates on the images of well-known points (reference<br />
marks of the bottom plate of the camera) and the modelling of the<br />
distortion by a bidimensional transformation. One usually chooses an<br />
affinity (general linear transformation) that represents correctly the main<br />
part of the distortion. But this distortion is thus measured only on the side<br />
of the image, in a limited number of points (eight in general). The interpolation<br />
inside the image is therefore of low reliability.<br />
1.1.3.6 Modelling of non-quantifiable defects<br />
The main corrections to apply to the simplistic model of the central perspective<br />
have been reviewed. Several of them can be considered as perfectly<br />
known – the curvature of the Earth for example; others can be calculated
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
roughly: refraction, distortion of the film, and others that are completely<br />
unknown, the distortion of the porthole glass plate, for example.<br />
When the precise geometric reconstitution is indispensable – as is notably<br />
the case in aerotriangulation, where the addition of systematic errors can<br />
generate intolerable imprecision – a counter-measure consists in adding to<br />
the equation of collinearity and its corrective terms a parametric model<br />
of distortion (often polynomial). This model contains a small number of<br />
parameters, chosen in order to best represent influences of no directly<br />
modelizable defects. These supplementary a posteriori unknowns of systematism<br />
will be solved simultaneously with the main unknowns of the<br />
photogrammetric system, which are coordinates of the ground points, tops<br />
of images, and rotations of images. This technique usually permits a gain<br />
of 30 to 40 per cent on altimetric precision of the aerotriangulation.<br />
1.1.4 Some convenient rules<br />
The geometry of the aerial image 15<br />
Rules for the correct use of aerial photogrammetry have been known for<br />
several decades, and we will therefore only briefly recall the main features.<br />
Besides, variables on which one can intervene are not numerous. This<br />
arises essentially from the necessary precision, which can be specified in<br />
the requirements. They can also be deduced from the scale of the survey<br />
(even though this one is digital, one can assign it a scale corresponding to<br />
its precision, that is conventionally 0.1 mm to the scale). The practice<br />
shows that the errors that accumulate during the photogrammetric process<br />
limit its precision – in the current studies – to about 15 to 20 microns on<br />
the image (on the very well-defined points) for each of the two planimetric<br />
coordinates. This number can be improved appreciably (down to 2 to<br />
3 m) by choosing some more coercive operative (and simultaneously<br />
costlier) ways to work. For the altimetry, this number must be grosso<br />
modo divided by the ratio between the basis (distance between points of<br />
view) and the distance (of these points of view to the photographed object).<br />
This ratio depends directly on the chosen field angle for the camera: the<br />
wider the field, the more this ratio will increase, the more the altimetric<br />
precision will be close to the planimetric precision. As an example, Table<br />
1.1.5 sums up the characteristics of the most current aerial camera values<br />
(in format 24 × 24 cm) with values of standard base length (overlap of<br />
55 per cent).<br />
Knowing the precision requested for the survey, it is then very simple<br />
to determine the scale of the images. The choice of the focal distance will<br />
be guided by the following considerations: the longer the focal distance,<br />
the higher the altitude of flight will be, the more the hidden parts will be<br />
reduced (essentially feet of buildings in city, not to mention complete<br />
streets), but simultaneously the more the altimetric precision is degraded,<br />
and the more marked the radiometric influence of the atmosphere.
16 Yves Egels<br />
Table 1.1.5 Characteristics of the most current aerial camera values<br />
(in format 24 × 24 cm) with values of standard baselength<br />
(overlap of 55 per cent)<br />
Focal length 88 mm 152 mm 210 mm 300 mm<br />
Field of view 120° 90° 73° 55°<br />
B/H 1.3 0.66 0.48 0.33<br />
Flight height (1/10,000) 880 m 1520 m 2100 m 3000 m<br />
Planimetric precision (1/10,000) 20 cm 20 cm 20 cm 20 cm<br />
Altimetric precision (1/10,000) 15 cm 30 cm 40 cm 60 cm<br />
In practice, one will preferentially use the focal distance of 152 mm.<br />
Shorter focal lengths will be reserved for the very small scales, for which<br />
the altitude of flight can constitute a limitation (the gain in altimetric precision<br />
is actually quite theoretical, as the mediocre quality of the optics<br />
makes one lose what was gained in geometry). A long focal distance will<br />
be preferred most of the time for the orthophotographic surveys in urban<br />
zones, in which the loss of altimetric quality is not a major handicap. In<br />
the case of a colour photograph, it will be necessary nevertheless to take<br />
care of the increase of the atmospheric diffusion fog (due to a thicker<br />
diffusing layer), which will especially have the consequence of limiting the<br />
number of days when it is possible to take the images. Another way (less<br />
economic) to palliate the hidden part problem is to use a camera with a<br />
normal focal distance, while choosing more important overlaps: one can<br />
reconcile a good altimetric precision, a weak atmospheric fog, and acceptable<br />
distortions for objects on the ground.<br />
1.2 RADIOMETRIC EFFECTS OF THE ATMOSPHERE<br />
AND THE OPTICS<br />
Michel Kasser<br />
1.2.1 Atmospheric diffusion: Rayleigh and Mie diffusions<br />
The atmosphere is generally opaque to the very short wavelengths, its<br />
transmission only starting toward 0.35 m. Then, the atmosphere presents<br />
several ‘windows’ until 14 m, the absorption becoming again practically<br />
total between 14 m and 1 mm. Finally, the transmission reappears<br />
progressively between 1 mm and 5 cm, to become practically perfect to<br />
the longer wavelengths (radio waves).<br />
The low atmosphere, in the visible band, is not perfectly transparent,<br />
and when it is illuminated by the Sun it distributes a certain part of the
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Radiometric effects of the atmosphere, optics 17<br />
radiance that it receives: all of us can observe the milky white colour of<br />
the sky close to the horizon, and to the contrary its blue colour is very<br />
marked in high altitude sites towards the zenith. These luminous emissions<br />
are important data to understand correctly the images used in<br />
photogrammetry, and in particular the diffuse lighting coming from the<br />
sky in zones of shades, or again the atmospheric fog or white-out and its<br />
variation according to the altitude of image acquisition. We will start with<br />
a presentation of the present diffusion mechanisms in the atmosphere, to<br />
assess the effects of this diffusion on the pictures obtained from an airplane<br />
or a satellite.<br />
1.2.1.1 Rayleigh diffusion<br />
The diffusion by the atmospheric gases (often called absorption, in an erroneous<br />
way, since the luminous energy removed from the light rays is merely<br />
redistributed in all directions, but not transformed into heat) is due to the<br />
electronic transitions or vibrations of atoms and molecules generally present<br />
in the atmosphere. On the whole, the electronic transitions of atoms<br />
provoke a diffusion generally situated in the UV (these resonances are very<br />
much damped because of the important couplings between atoms in every<br />
molecule, and thus the effect is still noticeable even at much lower frequencies),<br />
whereas vibration transitions of molecules rather provoke diffusions<br />
situated at lower frequencies, and therefore in the IR (and these are sharp<br />
resonances, because there is nearly no coupling between the elementary<br />
resonators that are molecules, and therefore the flanks of bands are very<br />
stiff). The Rayleigh diffusion, particularly important to the high optical<br />
frequencies, corresponds to these atomic resonances. We have therefore a<br />
factor of attenuation of a light ray when it crosses the atmosphere, and<br />
at the same time a substantial source of parasitic light, since light removed<br />
from the incidental ray is redistributed in all directions, so that the atmosphere<br />
becomes a secondary light source.<br />
In short, Rayleigh diffusion is characterized by two aspects:<br />
• its efficiency varies as 4 , which combined with the spectral sensitivity<br />
of the human eye is responsible of the blue appearance of the sky;<br />
• the diagram of light redistributed is symmetrical between the before<br />
and the rear, with a front or rear diffusion twice as important as the<br />
lateral one.<br />
This diffusion may thus be very satisfactorily and accurately modelled.<br />
1.2.1.2 Diffusion by sprays (Mie diffusion)<br />
One calls sprays all particles, droplets of water, dusts that are continuously<br />
sent into the atmosphere by wind sweeping soil, but that can come also
18 Michel Kasser<br />
Table 1.2.1 Typical particle size in<br />
various sprays<br />
Smoke 0.001–0.5 m<br />
Industrial smoke 0.5–50 m<br />
Mist, dust 0.001–0.5 m<br />
Fog, cloud 2–30 m<br />
Drizzle 50–200 m<br />
Rain 200–2000 m<br />
from the sea (particles of salt), from volcanoes because of cataclysmic explosions,<br />
from human activities (industrial pollution, cars, forest fires, etc.) and<br />
also from outside of our globe: meteoroids and their remnants. The essential<br />
characteristic of these sprays, since such is the generic name that one<br />
usually gives to all of these non-gaseous components of the atmosphere, is<br />
that their distribution is at any instance very variable with the altitude (presenting<br />
a very accentuated maximum to the neighbourhood of the ground)<br />
but also very variable with the site and with time. (See Table 1.2.1.)<br />
In fact, every spray distributes light, in the same way that gaseous<br />
molecules do. However, an important difference is that molecules present<br />
some weaker dimensions than the wavelength of light that interests us<br />
here. This implies that the Rayleigh diffusion by molecules (or atoms)<br />
is more or less isotropic. To the contrary, the diffusion by sprays, whose<br />
dimensions are often the same order as the wavelengths considered,<br />
present strong variations with the direction of observation. At the same<br />
time, the intensity of the diffusion by sprays – also called the Mie<br />
diffusion, depends strongly on the dimension of the diffusing particle.<br />
The exact calculation is of limited value, since in practice the quantity of<br />
sprays in the atmosphere is very variable according to the altitude, and<br />
also especially according to the time. In any case, it is generally not possible<br />
to foresee it.<br />
In practice, it is to the contrary from luminous diffusion measurements<br />
that one may measure the density (in number of particles/cm 3 ) of sprays<br />
according to the altitude.<br />
Table 1.2.2 provides some figures, but it is necessary to note that<br />
this distribution is reasonably steady only to an altitude over 5 km. In<br />
the first kilometres of the atmosphere, to the contrary, the influence of<br />
wind, of human activities is such that numbers in the table are only indicative.<br />
The Mie diffusion is characterized by a directional diagram that varies<br />
markedly with the size of sprays. For the very small particles compared<br />
to the wavelength, the diagram is very similar to that of Rayleigh diffusion.<br />
For larger particles, it becomes more and more dissymetric between front
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Table 1.2.2 Typical spray density vs.<br />
altitude<br />
h Sprays density<br />
(km) (particles per cm 3 )<br />
0 200<br />
1 87<br />
2 38<br />
3 16<br />
4 7.2<br />
5 3.1<br />
6 1.1<br />
7 0.4<br />
8 0.14<br />
9 0.05<br />
Radiometric effects of the atmosphere, optics 19<br />
and rear, to give a retrodiffusion negligible compared to the importance<br />
of the front diffusion. (See Figure 1.2.3.)<br />
To assess the importance of the phenomenon on aerial image acquisitions,<br />
one determines the attenuation of a light ray travelling upwards,<br />
from the observed horizontal attenuation, which is obviously more accessible.<br />
One characterizes it by the ‘meteorological visibility’ V, that is the<br />
distance to which the intensity of a collimated light ray of 0.555 m<br />
(wavelength of the maximum sensitivity the eye) falls to 2 per cent of its<br />
initial value. Under these conditions, the practice shows that the diffusion<br />
that comes with the attenuation removes all observable details of the<br />
observed zone, that melts in a uniform grey colour.<br />
R = 0.05 µm, scale × 0.003<br />
θ<br />
R = 0.1 µm, scale × 3<br />
R = 0.5 µm, scale × 1000<br />
Figure 1.2.3 Angular distribution of the Mie diffusion for a wavelength of<br />
0.5 m and for three particle sizes: 0.05, 0.1 and 0.5 m.
20 Michel Kasser<br />
The reduction of intensity of a light ray depends, physically, on the<br />
number of particles interacting with it, the resulting attenuation being<br />
correctly described by Beer’s law:<br />
I I 0 e kx , (1.28)<br />
where x is the distance in the absorbing medium, and k the coefficient of<br />
absorption.<br />
The visibility V, easily appreciated by a trained observer, allows the<br />
calculation K M K R and therefore the part of attenuation due to sprays,<br />
by the difference between the observed coefficient K obs (such as: e K obsV <br />
0.02) and the one K R easy to calculate, resulting from Rayleigh diffusion.<br />
I I 0 e (K R K M ) x I 0 e K R x e K M x .<br />
(1.29)<br />
From the visibility V, one establishes in an experimental way the coefficient<br />
of attenuation for different wavelengths. This leads to the formula<br />
(in km 1 , for in m and V in km):<br />
K M 3.91<br />
V 0.555<br />
0.585V1/3<br />
. (1.30)<br />
Table 1.2.4 gives the numeric figures for various wavelengths.<br />
If one wants to switch from the horizontal attenuation to the vertical<br />
attenuation, to assess the effect on a light ray going from the ground to<br />
a plane, one admits besides that the density of sprays, according to the<br />
altitude, follows an exponential law. Figure 1.2.5 shows the relation<br />
between the coefficient of extinction (that is the k coefficient of Beer’s law)<br />
and the visibility.<br />
Figure 1.2.6 illustrates the result of the integration calculation for a vertical<br />
crossing of the diffusing layer, done for an attenuation K M 10 3 km 1 at<br />
h 5 km (altitude over which the concentration of sprays stops being influenced<br />
considerably by ground effects, and may be considered as constant).<br />
Table 1.2.4 Variations of the Mie and Rayleigh coefficients of attenuation<br />
at various wavelengths for three values of the ground<br />
meteorological visibility<br />
(m) K Rayleigh K Mie<br />
(km 1 ) (km 1 )<br />
V 1 km V 5 km V 10 km<br />
0.4 0.044 4.7 1.1 0.58<br />
0.6 0.0083 3.7 0.72 0.35<br />
0.8 0.0026 3.1 0.54 0.24<br />
1.0 0.0006 2.8 0.43 0.18
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Altitude (km)<br />
7<br />
5<br />
3<br />
1<br />
V = 10 km<br />
V = 6 km<br />
V = 2 km<br />
Radiometric effects of the atmosphere, optics 21<br />
0.01 0.1 1<br />
KM coefficient at 0.555 µm in km –1<br />
Figure 1.2.5 Variations of the coefficient of attenuation at 0.555 m for three<br />
values of the ground meteorological visibility.<br />
Optical<br />
transmission in %<br />
100<br />
80<br />
1<br />
60<br />
0.6<br />
40<br />
20<br />
0.2<br />
10 0.1<br />
2 4 6 8 10 12(km)<br />
Meteorological visibility at ground level<br />
One should also note that, in the visible domain, the content of water<br />
vapour does not change the visibility.<br />
1.2.2 Effects of the atmospheric diffusion in aerial image<br />
acquisition<br />
The diffusion of light by atmospheric gas molecules and by all sorts of<br />
sprays dispersed in the atmosphere is not only responsible for the light<br />
attenuation. In fact, the portion of atmosphere crossed by the light rays<br />
of observation behaves, when it is illuminated by the Sun, like a luminous<br />
source whose radiance arrives in part on the detector.<br />
2<br />
Total optical density<br />
through the layer<br />
of concentration<br />
of sprays (0–5 km)<br />
Figure 1.2.6 Optical transmission and optical thickness for a vertical crossing of<br />
the diffusing layers for various values of meteorological visibility at<br />
ground level.
22 Michel Kasser<br />
Being foreign to the ground that one generally intends to observe, this<br />
radiance can be qualified as ‘parasitic’, and constitutes one of the main<br />
sources of ‘noise’ in image processing in the visible domain. This diffusion<br />
is as difficult to calculate as the corresponding attenuation, as like it,<br />
it depends on distribution by sizes and by the nature of diffusing particles.<br />
And moreover it depends, in the case of sprays whose size is not<br />
negligible in comparison to , on the angle between the direction of observation<br />
and the direction of lighting. The contribution of the Rayleigh<br />
diffusion, especially important to the short wavelengths, is therefore the<br />
only one easy to calculate.<br />
It remains to determine the fraction of this flux sent back towards the<br />
instrument of observation. To a first approximation, one will be able to<br />
admit that the diffusion is isotropic. The diffusion (sum of contributing<br />
diffusions Mie and Rayleigh) intervenes finally in the three following aspects:<br />
• Lighting brought by the sky in zones to the shade, diffuse source that<br />
includes at least the Rayleigh diffusion, and that is therefore more<br />
important in the blue, but that is in any case characteristic of the shade<br />
contrast. If this contrast is strong, it means that shades are poorly illuminated,<br />
and therefore that the diffusion is weak. The lower the Sun<br />
is on the horizon, the weaker the direct solar lighting and the more<br />
the atmospheric diffusion becomes predominant. On an indicative<br />
basis, between a surface exposed to the Sun and the same surface in<br />
the shade, the typical lighting differences are of: 2 (Sun to 20° on the<br />
horizon, sky misty enough, V 2 km), 3 (Sun to 45°, V 5 km),<br />
10 (extreme case, very high Sun and V > 10 km, frequent case, for<br />
example, in high mountains).<br />
• Attenuation of the useful light ray going from the illuminated ground<br />
towards the device used for image acquisitions. This attenuation<br />
depends on the height of flight H and the value of V. For example<br />
for H > 4 km, the attenuation is the same as that of the direct solar<br />
radiance, and it may range from a value of transmission of 0.90 (V<br />
10 km, for example) to 0.70 (V 3 km) or even less. And for the<br />
lower flight heights, the exponential decrease of K with the increasing<br />
altitudes shows that this coefficient is near unity for the low flights<br />
(large scales, H 1,500 to 2,000 m). On the other hand, for values<br />
of H > 4 or 5 km, the attenuation is the same as for an imaging satellite.<br />
• Superposition of a reasonably uniform atmospheric fog or white-out<br />
on the useful picture formed in the focal plane. This parasitic luminous<br />
flux can be computed according to the following way in the case<br />
where the plane is entirely over the diffusing layer (H > 4 or 5 km):<br />
(1) the diffusion is reasonably symmetrical fore/rear; (2) this diffusion<br />
essentially originates from the light removed from the direct solar rays,<br />
between 10 per cent and 50 per cent according to our previous remarks;
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Radiometric effects of the atmosphere, optics 23<br />
(3) the half of this flux going upwards therefore represents the source<br />
of light, coming from all the lower half-space, the other half going<br />
downwards; (4) one makes the hypothesis that this flux is isotropic in<br />
all this half-space and one calculates, considering the aperture of the<br />
camera, the flux F that enters there from all this half-space; (5) one<br />
calculates the solid angle under which is seen a pixel on the ground<br />
by the camera; (6) one finally corrects F of the ratio formed by this<br />
solid angle and 2 (solid angle formed by a half-space). This figure<br />
for the flux will be reduced if the plane flies lower, as mentioned previously,<br />
but one is generally surprised when one does this calculation,<br />
as the atmospheric fog is so important as compared to the useful signal.<br />
In current conditions, for H > 5 km the energy from the atmospheric<br />
fog is frequently more than the energy of the useful image, which<br />
almost always gives to the uncorrected colour pictures a milky white<br />
aspect that makes the appearance of colours often disappear to a large<br />
extent.<br />
1.2.3 Radiometric problems due to the optics<br />
In the optic energy calculations that one does in aerial imagery, it is necessary<br />
to also take into account the classic effects due to camera optics. We<br />
have been used, with traditional aerial cameras, to seeing these defects<br />
corrected in a quasi-perfect way, essentially because the downstream photogrammetric<br />
chain was unable to correct any radiometric defect of the<br />
optics. The digital image acquisitions using a CCD permit one to deal with<br />
these problems very correctly even with much cheaper and readily available<br />
materials, optimized for the quality of the picture, and not for typically<br />
photogrammetric aspects such as the distortion: these shortcomings are<br />
corrected by calculation. Remarks that follow sum up a few of these points:<br />
• The field curvature is very inopportune, since it implies that the zone<br />
of better focusing is not plane: this defect must therefore be further<br />
reduced if one works with small pixels in the picture plane. For that<br />
reason it is necessary to pay attention to the parallel glass plates in<br />
front of the CCD sensor when one works with a large field, which is<br />
the rule in photogrammetry: this plate leads to an easily observable<br />
radial distortion, although not a troublesome one since it can be<br />
included in the calibration.<br />
• The distortion has inevitably a symmetry of revolution. It is often<br />
described by an odd degree polynomial, and one should not be<br />
concerned by its importance, but rather by the residues observed<br />
between the polynomial and measures of calibration.<br />
• The vignettage is governed by optics in the approximation of Gauss<br />
of a quite appreciable lighting loss of the sensitive surface following<br />
a law in cos 4 of the angle between rays and the optic axis, which
24 Michel Kasser<br />
reduces the brightness greatly within the angles of an image. This effect<br />
is very well corrected today, including optics for the general public,<br />
and can also be radiometrically calibrated in any camera, using an<br />
integrating sphere as a light source.<br />
• The shutter is also an important piece. If it is installed close to the<br />
optic centre of the objective, its rise and fall times do not create a<br />
problem, as at each step of the shutter movement, all the focal plane<br />
receives the same light modulation, but then there is a shutter in each<br />
objective used, which increases their cost. For the digital cameras using<br />
some CCD matrixes, one will note that the very high sensitivity of<br />
these devices requires having access to a very short exposure time,<br />
which is not necessarily feasible in times of strong sunlight, considering<br />
the relative slowness of these mechanical devices. One is then<br />
sometimes obliged to add neutral density filters in front of the objective<br />
to limit the received flux, as it is impossible to effectively close<br />
the diaphragm, which would create excessive diffraction effects.<br />
Otherwise, if one uses a linear movement blade shutter with this type<br />
of camera, it is necessary to check that the scrolling of the window<br />
proceeds in the sense of the transfer of charges of the CCD. And the<br />
electronic compensation of forward motion (against linear blurring<br />
due to the speed of the airplane, often called TDI – see §1.5, that<br />
consists in shifting charges in the focal plane at the same physical<br />
speed as the scrolling of the picture in the focal plane) has to be activated<br />
during the entire length of movement of the blades (generally<br />
1/125th of a second), and not during the time of local exposure of<br />
the CCD (that can range sometimes up to the 1/2,000th of a second).<br />
If the blade movement were oriented in another direction, it would<br />
indeed result in an important distortion of the image. And if its direction<br />
is good but the electronic compensation of forward movement is<br />
not activated at the correct time, there will only be a part of the image<br />
that will be correct, in sharpness and in exposure as well.<br />
• Finally, on digital cameras, it is necessary to note that it is easy to<br />
calibrate the sensitivity of pixels individually, and to apply a systematic<br />
correction, which improves the linearity of the measure. Otherwise,<br />
in the matrix systems, the electronic compensation of forward motion<br />
leads to sharing of the total response over several pixels, which<br />
improves the homogeneity of the pixel response once more.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
1.3 COLORIMETRY<br />
Yannick Bergia, Michel Kasser<br />
1.3.1 Introduction<br />
The word ‘colour’ is a dangerous word because it has very different senses.<br />
We all have one intuitive understanding of this notion, and we think we<br />
know what is green, or orange. In fact, the word ‘colour’ regains two<br />
fundamentally different notions:<br />
• on the one hand, the visual sensation that our eyes receive from the<br />
surface of an illuminated object;<br />
• on the other hand, the spectral curve of a radiance fixing, for every<br />
wavelength, the quantity of light emitted or transmitted by an object.<br />
The first meaning is eminently subjective and dependent on many parameters.<br />
The second is entirely objective: it corresponds to a physical<br />
measure. The fact that the language only proposes one term for these two<br />
notions is a source of confusion. We must always have to mind that our<br />
eyes only communicate to us a transcription of discerned light limiting<br />
itself to the supply of three values, and are not therefore capable of<br />
analysing spectral curves by a large number of points.<br />
The objective of this presentation is to attract the reader’s attention to<br />
the complexity of situations that will be met when some digital processes<br />
are to be applied, not to black and white pictures, but to pictures in colour.<br />
It will especially be the case, in digital photogrammetry, in processes of<br />
picture compression and automatic correlation. It is indeed certain that<br />
for such processes, it has been necessary since the conception of algorithms<br />
to take into account a given type of space colorimetry, and that it is necessary<br />
to choose it in functions, for example, of metrics problems that it<br />
presents.<br />
Colours that our eyes are capable of differentiating, if they are not in<br />
infinite number, are nevertheless extremely numerous. To compare them<br />
or to measure them, one refers to the theoretical basic colours (often called<br />
primary) that are defined as follows. One divides the visible spectrum into<br />
three sub domains, for which one defines colours with a constant radiance<br />
level in all spectral bands:<br />
• between 0.4 and 0.5 microns, the blue;<br />
• between 0.5 and 0.6 microns, the green;<br />
• between 0.6 and 0.7 microns, the red.<br />
Colorimetry 25<br />
A light is said to be blue therefore (respectively, green, red) if it is composed<br />
of equal parts of radiances of wavelengths between 0.4 and 0.5 microns
26 Yannick Bergia and Michel Kasser<br />
(respectively 0.5/0.6, 0.6/0.7 m) and does not contain any other wavelength<br />
of the visible spectrum (one does not consider the remainder of the<br />
spectrum). One also defines, next to these ‘fundamental’ colours, basic<br />
colours called ‘complementary’:<br />
• cyan, whose wavelengths range from 0.4 to 0.6 m;<br />
• magenta, from 0.4 to 0.5 m and 0.6 to 0.7 m;<br />
• yellow, from 0.5 to 0.7 m.<br />
1.3.2 Trichromatic vision<br />
The notion of trichromatic vision is one of the principles basic to<br />
colorimetry. It is based on the fact that the human eye possesses three<br />
types of receptors (cones) sensitive to colour (the eye also possesses other<br />
receptors, short sticks that are sensitive to brightness), each of the three<br />
types of cones having a different spectral sensitivity. This principle applies<br />
in such a way that a colour can be represented quantitatively by only three<br />
values. Besides, experiences prove that three quantities at a time are necessary<br />
and sufficient to specify a colour. Also, they introduce notions of<br />
additivity and proportionality of colours and so confer the intrinsic properties<br />
of a vectorial space of dimension three to all spaces where one would<br />
try to represent colours.<br />
1.3.3 Modelling of colours<br />
To specify a colour means to associate with the spectrum of distribution<br />
of energy of a given light a triplet of values that will represent it in what<br />
will be then a colorimetric space. Considering the results of experiences,<br />
one can consider that providing three colours respecting conditions of<br />
mutual independence (and that will be then the three primary axes of the<br />
space) can be sufficient to define a colorimetric space: for a given colour,<br />
multipliers coefficients of the three intensities of sources associated to<br />
primaries (when the reconstituted colour corresponds to the considered<br />
colour) will be coordinates of the colour in the constructed space. Another<br />
way to express this is to return to the physical data that contains the<br />
colour information and to integrate the spectral distribution of optical<br />
power on the visible wavelengths while using three functions linearly independent<br />
of weighting to define three channels. Thus, one can construct a<br />
colorimetric space, for example the fictive ABC system, by defining three<br />
functions of weighting a(), b(), c() (see Figure 1.3.1). Then every transformations,<br />
linear or not, of such a colorimetric space will also be a<br />
colorimetric space. There are an infinite number of such spaces, but one<br />
will find – in what follows – only certain spaces are often met with.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
a b c<br />
400 700<br />
1.3.4 Survey of classic colorimetric spaces, the RGB, the<br />
HSV and the L*a*b*<br />
1.3.4.1 The RGB space<br />
This is the most often met colorimetric space, used in particular for television<br />
and computer screens, of so-called ‘additive synthesis’ (as opposed<br />
to the ‘subtractive synthesis’ while using white light and subtracting from<br />
it a part of the spectra by filters, which is for example the case in printing,<br />
where each ink acts like a filter on the white paper). Its definition rests<br />
on the choice of three primary colours: the red, the green and the blue.<br />
Considering data processing, the RGB space is represented as a<br />
discretized cube of 255 units of side. All feasible colours for computer<br />
applications are represented therefore by a triplet (R, G, B) of the discrete<br />
space. We may note the first diagonal of the reference frame where colour<br />
is distributed with the same value for the three channels, R, G and B:<br />
it is the axis of the grey or achromatic axis.<br />
This space does not present specific points, but it is the one that the<br />
current material environment imposes on us: the large majority of pictures<br />
in true colours susceptible of being processed is coded in this system. We<br />
will also present methods of conversion from RGB towards another very<br />
common system, the HSV space, this only as an example.<br />
1.3.4.2 The HSV space<br />
⎧<br />
⎪A<br />
=<br />
⎪<br />
⎪<br />
⎨B<br />
=<br />
⎪<br />
⎪<br />
⎪C<br />
=<br />
⎪⎩<br />
Figure 1.3.1 Definition of a fictive colorimetric reference.<br />
Colorimetry 27<br />
a(<br />
λ ) ⋅ P(<br />
λ ) dλ<br />
b(<br />
λ ) ⋅ P(<br />
λ ) dλ<br />
c(<br />
λ ) ⋅ P(<br />
λ ) dλ<br />
Definitions<br />
The Hue, the Saturation and intensity Value that constitute the three<br />
components of this system are the relatively intuitive notions that allow a<br />
better description of the experience of the colour. Some more theoretical<br />
definitions can be given.<br />
λ<br />
700<br />
∫<br />
400<br />
700<br />
∫<br />
400<br />
700<br />
∫<br />
400
28 Yannick Bergia and Michel Kasser<br />
Cyan<br />
Green<br />
Blue<br />
Hue<br />
Yellow<br />
Magenta<br />
Red (Hue = 0)<br />
Figure 1.3.2 Distribution of basis colours on the circle of hues.<br />
The intensity can be presented as a linear measure of the light power<br />
emitted by the object that produces the colour. It allows the dark colours<br />
to be distinguished naturally from the bright ones.<br />
The hue is defined by the CIE (Commission Internationale de l’Éclairage)<br />
as the attribute of the visual perception according to which a zone appears<br />
to be similar to a colour among the red, the yellow, the green and the<br />
blue or to a combination of two of them. It is a way of comparing known<br />
colours on which it is easy to agree. According to this data, colours can<br />
be represented on a circle. The red, the green and the blue are there with<br />
an equal repartition (the red having by convention a null hue). Figure 1.3.2<br />
illustrates this distribution.<br />
The saturation allows the purity of a colour to be determined. It is while<br />
making the saturation vary from a null hue colour, for example, that one<br />
will move from the pure red (most saturated), that can be placed on the<br />
circumference of the hue circle, toward the neutral grey in the centre while<br />
passing all pink tones along the radius of the disk. From a more physical<br />
point of view one can say that if the dominant wavelength of a spectrum<br />
of power distribution defines the hue of the colour, then, the more this<br />
spectrum will be centred around this wavelength, the more the colour will<br />
be saturated.<br />
Considering these definitions, colours specified in the HSV space can be<br />
represented in a cylindrical space and are distributed in a double cone (see<br />
Figure 1.3.3).<br />
Distance in the HSV space<br />
To be able to exploit this space for process applications it is necessary<br />
to define a distance. Naturally, the Euclidian distance does not have any<br />
sense here, the obvious non-linearity of its cylindrical topology and in<br />
particular the discontinuities met in the neighbourhood of the axis of inten-
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
t<br />
i<br />
White<br />
s<br />
Black<br />
Red<br />
Figure 1.3.3 Theoretical representation of the HSV space.<br />
sity and the passage of the null hue are unacceptable. One has thus agreed<br />
to consider as a distance between two points in this duplicate cone the<br />
length of the smallest bow of a regular helix that it is possible to draw<br />
between them.<br />
Let us consider two points of the HSV space therefore, P 1 (I 1, T 1, S 1)<br />
and P 2 (I 2 , T 2 , S 2 ). One stands in an orthogonal Euclidian reference frame<br />
whose vertical axis (Z) is confounded with the axis of intensities, and the<br />
first axis (X) is directed by the vector of hue of the point whose intensity<br />
is smallest (suppose that it is P 1 ). One notes I the positive difference of<br />
intensity between the two points (in our case I I 2 I 1 ), S the difference<br />
of saturation (S S 2 S 1 ) and the smallest angular sector defined<br />
by the two angles of hues.<br />
We are then in the situation illustrated by Figure 1.3.4 showing the bow<br />
of helix for which we need to calculate the length precisely. In these conditions<br />
we propose the following parametric representation of this bow:<br />
P(t) X(t) S 1 t S) cos (t)<br />
Y(t) (S 1 t S) sin (t)<br />
Z(t) t I<br />
Colorimetry 29
30 Yannick Bergia and Michel Kasser<br />
Y<br />
Z<br />
θ<br />
The length of the bow is then given by the calculation of the curvilinear<br />
integral<br />
that provides the following result:<br />
D h <br />
S 1<br />
1<br />
Dh 0<br />
S 2<br />
|| ∂P(t)<br />
∂t || dt<br />
1<br />
2 S S 2 √(S 2 I 2 S 2 2 ) S 1 √(S 2 I 2 S 1 2 ) <br />
S 2 I 2<br />
<br />
P 1<br />
P 2<br />
P(t)<br />
argsh <br />
X<br />
Figure 1.3.4 Portion of helix to compute distances in the HSV colorimetric<br />
system.<br />
S<br />
argsh<br />
1<br />
<br />
.<br />
√(S2 I2 ) <br />
S2 √(S2 I2 <br />
)<br />
(1.31)<br />
Technical way of use of the HSV space<br />
There does not exist a unique HSV space strictly speaking, rather numerous<br />
definitions or derivative specifications. Indeed several quantities deduced<br />
for example from the R, G and B values of a specified colour in the RGB<br />
space can correspond to very subjective definitions of hue, intensity and<br />
saturation that have been given. Conceptually all these spaces are very<br />
much the same in the sense that they all allow one to easily specify a<br />
sensation of colour in terms of hue, intensity and saturation, or if need<br />
be neighbours and equivalent notions.<br />
The most often used definition is that found for example in Gonzales<br />
and Woods (1993): the calculation takes place on the r, v and b values<br />
between 0 and 1 (it is about the R, V and B initial values divided by 255).<br />
If r v b, the considered colour is a nuance of grey, and one imposes<br />
Z<br />
Y<br />
S 1<br />
P 1<br />
P 2<br />
P(t)<br />
X
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
by convention: t s 0; it does not affect the reversible character of the<br />
transformation.<br />
(r v) (r b)<br />
2 arc cos 2√(r2<br />
v2 b2 (r v) (r b)<br />
arc cos 2√(r<br />
, if b v<br />
rv vb br) 2 v2 b2 r v b<br />
i ,<br />
3<br />
<br />
, if v b<br />
rv vb br)<br />
t (1.32)<br />
s 1 <br />
min {r, v, b}<br />
i<br />
.<br />
Colorimetry 31<br />
1.3.4.3 The L*a*b* space<br />
L*a*b*, defined by the CIE in 1976, has the goal of providing a space of<br />
work in which the Euclidian distance has a true sense: it develops a space<br />
uniform to the perception, that is to say in which two greens and two<br />
blues separated by the same distance would lead to the same sensation of<br />
difference for the human eye, which will not be for example the case in<br />
y<br />
0.8<br />
λ = 520 nm<br />
λ = 380 nm<br />
λ = 700 nm<br />
0.8 µm<br />
Figure 1.3.5 Separation of colours in the chromatic L*a*b* diagram.<br />
x
32 Yannick Bergia and Michel Kasser<br />
the RGB system basis as we have a better capability to differentiate in the<br />
blues than in the greens. This affirmation has been verified by the experience<br />
of the chromatic matching of Wyszecki (Wyszecki and Stiles, 1982)<br />
whose results are presented on the chromatic diagram of Figure 1.3.5.<br />
Ellipses that are drawn there (whose size has been exaggerated nevertheless)<br />
represent, according to the position in the diagram, zones whose<br />
colours cannot be discerned by observers.<br />
The transformation toward L*a*b* will distort the space so that these<br />
ellipses become circles of constant radius on the whole of the diagram. It is<br />
this important property of linearity of perception that forms its essential<br />
interest: if one works in a metrics that replicates mechanisms of the human<br />
vision, one can expect that the result of segmentations obtained be relatively<br />
satisfactory, in the sense that it will correspond more to the delimitation that<br />
a human operator could have drawn by hand while seeing the picture.<br />
1.3.5 Distribution of colours of a picture in the different<br />
spaces<br />
In order to place this survey of colorimetric spaces in the context of the<br />
picture process, we have found it interesting to represent a picture in these<br />
different spaces. Naturally, whatever is the space in which the picture<br />
is coded, its aspect must remain the same, but not the distribution of its<br />
colours in the three-dimensional space of colours representation. One<br />
considers the aerial photograph of an urban zone as an example (see Figure<br />
1.3.6). And one gives the representation of the colour clouds that are<br />
present in several classic colorimetric spaces, including those seen previously<br />
(Figure 1.3.7).<br />
This horizontal comparison imposes an observation: if the cloud of points<br />
keeps, with the exception of the orientation, the same aspect in the two<br />
spaces linearly deriving from RGB (KL and XYZ), its geometry in the<br />
two uniforms spaces of the CIE (L*a*b* and L*u*v*) is radically different.<br />
The change of space has consequently a deep modification of the metrics<br />
Figure 1.3.6 Example of colour image used to test various colorimetric spaces,<br />
here displayed in black and white.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
that allow one to anticipate very variable results for the automatic process<br />
applications according to the chosen workspace.<br />
1.3.6 Conclusion<br />
Colorimetry 33<br />
Figure 1.3.7 Distribution of colours of the picture in the different colorimetric<br />
spaces.<br />
It is therefore important not to underestimate problems due to the use of<br />
colour in digital imagery, in particular for photogrammetric processes. At<br />
least one should know how to exploit the passage of a space to one dimension<br />
(the black and white) to a space of at least three dimensions. It is<br />
certain that in a few years there will be for the disposition of users some<br />
spatial images with small pixels of interest to photogrammetrists, and using<br />
more than three spectral bands. Processes used for the black and white<br />
will be reused with multi-spectral pictures, but it will be at the cost of a<br />
considerable loss of information. In fact it will be necessary to start from<br />
zero systematically the analyses of the studied phenomenon to extricate<br />
the best party possible of the available spectral bands. In this optics, one<br />
will need to adopt a colour space with a metrics adapted to the problem
34 Yannick Bergia and Michel Kasser<br />
each time: compression of the image, automatic correlation, automatic<br />
segmentation, automatic shape recognition, etc. Here we presented some<br />
details of trichromatic colorimetry, because it was necessarily very well<br />
studied for the human eye. It is necessary to note therefore that an equivalent<br />
study should be undertaken for a different number of spectral bands,<br />
at each time where the case will happen. . . .<br />
References<br />
Wyszecki G., Stiles W.S. (1982) Colour Science: Concepts Methods and Quantitative<br />
Dated Formulae, second edition, John Wiley & Sons.<br />
Gonzales R.C., Woods R.E. (1993) <strong>Digital</strong> Picture Processing, Addison-Wesley.<br />
1.4 GEOMETRY OF AERIAL AND SPATIAL PICTURES<br />
Michel Kasser<br />
The usable digital pictures in photogrammetry may originate from three<br />
possible sources: digitized traditional silver halide pictures, digital pictures<br />
originating from matrixes or airborne linear CCD cameras, and the spatial<br />
images of high and very high resolution (between 10 m and 0.8 m pixel<br />
size), these also constituted of CCD matrixes or linear CCD. The geometry<br />
of these different types of pictures is very different, and will be<br />
presented here according to their specificity.<br />
1.4.1 Images with a large field and images with a narrow<br />
field<br />
What distinguishes pictures of spatial origin from those using an airborne<br />
sensor is only the angular field of view, obviously narrower in the case of<br />
the first ones. This field has an influence on several aspects of images, in<br />
particular:<br />
• The direction of observation is pretty much constant in relation to the<br />
solar radiance for the spatial sensors, whereas it varies considerably<br />
in the case of pictures with a large field. Thus we have the evidence<br />
of a phenomenon of brilliant zone, called ‘hot spot’, which is due to<br />
the fact that when the direction of observation comes close to the<br />
direction of the solar radiance, one can no longer observe any zone<br />
of shade whatever the roughness of the ground, or the vegetation, or<br />
the buildings. This direction has a very weak probability to be met in<br />
an image with a narrow field. On the other hand one very often meets<br />
it in aerial pictures with a large field as soon as the sun is high on<br />
the horizon. It materializes a zone of the picture where the reflectance
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Geometry of aerial and spatial pictures 35<br />
of the ground becomes much higher than elsewhere. Another important<br />
point: the direction of reflection of the sun on surfaces of free<br />
water is, for the same reasons of size as for the field of vision of the<br />
sensor, susceptible of provoking a prompt blinding of sensors having<br />
a large field, a situation that is also statistically not frequent on the<br />
spatial high-resolution imaging devices.<br />
• The phenomena of refraction: the light rays captured by a spatial<br />
imager are practically all parallel (the angular field of the Ikonos satellite,<br />
launched in 1999, is 0.9°), and phenomena of differential refraction<br />
between the extreme rays are very weak, otherwise negligible. This<br />
refraction has only very few effects on the absolute pointing of the<br />
sensor in the space, since it leads to a curvature that is significant only<br />
in the last kilometres close to the ground, with therefore a negligible<br />
angular effect since the satellite is between 500 and 800 km away at<br />
least. These aspects are quite different for the airborne imagers with<br />
a large field of view (see §1.1).<br />
• On the other hand, for the phenomena bound to the atmospheric diffusion<br />
(see §1.2), there is no difference between an image obtained from<br />
an airborne sensor flying to an altitude higher than 4 or 5 km and an<br />
image obtained on board a satellite, since in the two cases all the<br />
diffusing layers have been crossed.<br />
• The stereoscopic acquisition on a plane is performed in a very simple<br />
way, essentially because the airplane can acquire any images only when<br />
it flies on a straight axis. According to the requested longitudinal<br />
overlap rate and to the speed of the plane in relation to the ground,<br />
one chooses a rate of image acquisition. And the axis of any image is<br />
in all cases practically vertical. On the other hand, on a satellite one<br />
can get a stereoscopy in several modes, along the track or between<br />
two different orbits. When one works in stereoscopy along the track<br />
(so-called ‘ahead–rear’ stereoscopy, between two similar instruments<br />
pointed ahead and behind, or with an instrument capable of aiming<br />
very quickly in the direction of the orbit), considering the speed of the<br />
satellite on its orbit (of the order of 7 km s 1 ), pictures are nearly<br />
exactly synchronous. In this case the direction of the stereoscopic base<br />
is by necessity that of the track of the satellite, that is imposed by<br />
laws of celestial mechanics and that is virtually impossible to change.<br />
But from several different orbits the satellite can also aim at the same<br />
scene, which provides a ‘lateral’ stereoscopy, and then a considerable<br />
time difference between acquisitions may happen, which sometimes<br />
complicates the work of photogrammetric restitution: for example the<br />
position of the sun is different and therefore shades changed a lot (this<br />
induces difficulties with the automatic correlators), vegetation changed<br />
(new leaves, or loss of leaves as well), cultures are completely different,<br />
levels of water in streams, lakes, or even the sea (tides) are no longer<br />
the same, vehicles have moved, etc. . . . to acquire pictures as quickly
36 Michel Kasser<br />
as possible, satellites are obliged to use the lateral pointing as well as<br />
the ahead–rear stereoscopy, but it is necessary to understand all the<br />
consequences of this for the acquired images when one considers<br />
photogrammetric processes.<br />
1.4.2 Geometry of images in various sensors<br />
There are essentially two main geometries of sensors that are used for<br />
photogrammetric processes: the conical geometry and the cylindro-conic<br />
geometry.<br />
1.4.2.1 Conical geometry: use of films or CCD matrixes<br />
This is well known, since it is the main means of human vision that has<br />
been used since the invention of photography by Nièpce or of the cinema<br />
by the Lumière brothers. It essentially implies a sequential mechanism to<br />
acquire a picture: a very brief exposure time, followed by one technically<br />
necessary dead time to restock a virgin length of film, or to transfer and<br />
to store the digital values corresponding to every pixel on a digital camera.<br />
This dead time is profitably taken on satellites to reorient the aiming system<br />
in order to image as many possible scenes in relation to the plan of work<br />
that is requested.<br />
If the necessary exposure time is too long, there occurs a phenomenon<br />
of linear blurring, because the picture moves continually in the focal plane<br />
with respect to the movement of the airplane or the satellite, which transforms<br />
the picture of a point into a small segment. To overcome this, the<br />
correction of linear blurring (forward motion compensation, or FMC) on<br />
cameras using silver halide films is performed by moving the film during<br />
the exposure time. This film is pushed, indeed, by depression on a very<br />
plane surface (the bottom plate of the camera), and a mechanism moves<br />
this plate in order to follow exactly the movement of the image in the<br />
focal plane during the short period of opening of the shutter. For the CCD<br />
matrixes it is even simpler; it is sufficient to provoke a charge transfer at<br />
the proper speed so that a given set of charges is created by only one<br />
elementary point of the ground: as the picture of this point moves, one<br />
will make the charges created move at the same speed. The FMC, in these<br />
two cases, requires absolute knowledge of the speed vector of the image<br />
in the focal plane, as well as a general mechanical orientation of the sensor<br />
so that its compensation capacity can be parallel to this speed vector. It<br />
is not necessarily easy to know this speed vector in an airplane, since the<br />
leeway of the plane under the lateral wind effect must be known also.<br />
Evidently these parameters of speed may be advantageously provided by<br />
the system of navigation of the plane, currently the GPS. And on a satellite,<br />
these parameters are deduced directly from the orbit and from<br />
coordinates of the target.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Geometry of aerial and spatial pictures 37<br />
1.4.2.2 Cylindro-conic geometry: use of a linear CCD<br />
This geometry is well known in remote sensing (airborne radiometers,<br />
Landsat satellites . . .), and has been used also in photogrammetry because<br />
of the Spot satellites without discontinuity since 1986. This geometry is<br />
provided naturally by every linear sensor (linear CCD) placed in the focal<br />
plane perpendicularly to the speed vector of the image. It is therefore the<br />
movement of the plane or the satellite that allows one of the dimensions<br />
of the image to be described. The geometric details of these sensors used<br />
in photogrammetry rely especially on the more or less stable speed vector.<br />
In a satellite, the trajectory is locally very precisely a Keplerian orbit,<br />
there is no lack of stability to fear on the parameters of position. On the<br />
other hand the satellite is frequently equipped with solar panels that<br />
are often reoriented automatically thanks to small motors, which create,<br />
in backlash, small changes of attitude. And generally it is impossible to<br />
guarantee a perfect aiming stability during the acquisition of an image,<br />
which results in a small distortion, that may be corrected in the most<br />
advanced means to process the geometry of the images.<br />
But on a plane, the small movements in roll, pitch and yaw are permanent,<br />
with an amplitude and a frequency that depend directly on the<br />
turbulence of air. Therefore inevitably one must perform a permanent record<br />
at least of the attitude, on the three axes, of the sensor. In fact one has also<br />
to measure permanently the three coordinates, as precisely as possible, of<br />
the optic centre of the sensor, using an inertial system and a precise GPS<br />
receiver whose data are quite complementary. Once the data are acquired<br />
one may reconstitute the images as one would have acquired them if the<br />
trajectory of the plane had been perfectly regular. But this correction, in<br />
spite of the efforts performed using excellent inertial platforms, does not<br />
succeed to perfectly correct these movements as soon as there is a little<br />
turbulence, which is very often the case for the urban surveys, generally<br />
done at a low altitude and therefore in a turbulent flight zone. This results<br />
in edges of buildings that are not straight, and finally residual image<br />
distortions that can often cover several pixels. In particular, there may be<br />
real hiatuses of data, owing to the angular movements of the plane being<br />
too rapid, and on another hand geometric aliasing of the picture, a unique<br />
point being imaged on several lines. Such discontinuities are sometimes<br />
impossible to correct by computation, and the only possible resource is then<br />
to ‘fill’ these hiatuses by interpolations between neighbouring strips, but<br />
this is not effectively very satisfactory in terms of picture quality. The only<br />
possibility to avoid such situations consists in installing the sensor on a<br />
servoed platform correcting in real-time movements of the plane, which is<br />
obviously a considerable overcost for the equipment.<br />
In counterpart, this technology permits data coming from several linear<br />
CCD to be combined, with potentially no limit to the number of pixels<br />
measured along a line.
38 Michel Kasser<br />
In this cylindro-conical geometry it is necessary to note a point that is<br />
absolutely fundamental in photogrammetric uses of the acquired images.<br />
Contrary to what occurs in a conical geometry, points are never seen on<br />
several consecutive pictures, since in fact every line of data is geometrically<br />
independent of the previous one. Thus the exact knowledge of the<br />
trajectory is mandatory to be able to put in correspondence the different<br />
successive groups of points seen by the linear CCD. The trajectography is<br />
therefore an essential element of geometry allowing the reconstitution of<br />
the relief, and the images alone are not sufficient to proceed to a photogrammetric<br />
restitution. The only possible validation of these parameters of<br />
trajectory, which become so important here, consists in using effectively<br />
three directions of observation in order to pull a certain redundancy of<br />
determination of the DTM (digital terrain model), that may validate the<br />
whole set of parameters introduced in the calculation. This point is fundamentally<br />
different from the usual geometry with the traditional cameras,<br />
since the stereoscopic basis doesn’t need to be known then (even if we<br />
find an appreciable economical interest in the use of a GPS receiver in the<br />
plane, and even an inertial platform too, in order to get the absolute orientation<br />
of the images, and thus to avoid the determination of ground<br />
reference points, always quite expensive to get; but this equipment is not<br />
mandatory), and that it is in any case redetermined by the calculation<br />
when the model is formed.<br />
Pictures obtained by such devices can be acquired in very long strips,<br />
continuously if need be, which suppresses worries of splices between images<br />
in the longitudinal sense, parallel to the axis of flight. The carving in<br />
successive images is therefore artificial and can be modified by the user<br />
without any problem. Here we should note also that a very strict calibration<br />
of the radiometric response of different pixels is mandatory, as it<br />
is for the linear CCD scanners used to digitize aerial pictures; any failure<br />
to complete it will result in unpleasant linear artefacts. For that reason<br />
the Ikonos satellite has included on board a calibration device regularly<br />
activated, that exposes all pixels to the solar light.<br />
As for the matrix systems, the stereoscopy can be used in the ‘ahead–rear’<br />
mode but also, as for satellites, by using lateral aiming (case of the Spot<br />
satellites). The airborne systems follow the principle of stereoscopy<br />
‘ahead–rear’ and are described in §1.5.2.<br />
The cylindro-conical geometry requires algorithms that are quite different<br />
from those used in classic conical geometry for photogrammetric restitution.<br />
These algorithms have been explored extensively for the restitution<br />
of the Spot images since the 1980s. Nevertheless, they remain poorly known<br />
and forbid de facto the use of equipment and software not designed for<br />
this geometry, which thus prevents the use of these data on other stereoplotting<br />
devices.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
1.4.3 Conclusion<br />
<strong>Digital</strong> image acquisition 39<br />
The different geometries used to acquire pictures have some important<br />
consequences when a photogrammetric process is to be performed. It is<br />
often not just a detail, since if one doesn’t take the necessary precautions<br />
one may find it impossible to validate by internal means the results<br />
obtained. Much attention must be paid to the practical consequences of<br />
choices of images, in particular with the arrival on the market of very high<br />
spatial resolution images, apparently very similar to the aerial pictures,<br />
but which have completely different geometric properties.<br />
In the domain of the large field airborne imagery, linear CCDs systems<br />
seem especially destined to provide very high resolution pictures, but whose<br />
metrics is not very critical (devoted to photo-interpretation, for example).<br />
On the other hand the matrix systems, providing an extremely rigorous<br />
geometry, will be much more suitable to photogrammetric uses.<br />
1.5 DIGITAL IMAGE ACQUISITION WITH AIRBORNE<br />
CCD CAMERAS<br />
Michel Kasser<br />
1.5.1 Introduction<br />
Currently most processes of photogrammetry turn toward a larger and<br />
larger use of digital images. Considering the reliability and the quality of<br />
materials for aerial image acquisition used several decades ago, but also<br />
considering the excellent knowledge that we have about the defects of this<br />
chain of image acquisition, research laboratories and constructors wisely<br />
started with the reuse of previous achievements. They proposed therefore<br />
first to digitize images obtained in aerial image acquisitions on argentic<br />
traditional film. Nevertheless, this solution is only a temporary solution,<br />
and as we will see in §1.8, the digitization of films is a delicate and costly<br />
operation, which brings its own share of deterioration to images. It is<br />
therefore natural that studies have been led, initially, since 1980, at the<br />
DLR in Germany for the Martian exploration, then in France since 1990,<br />
to get digital images directly in the airplane itself. Studies moved in two<br />
simultaneous directions, using two CCD devices: linear CCD, according<br />
to the model of sensor popularized by the spatial imagers like SPOT or<br />
Ikonos, using the advancement of the plane to scan one direction of the<br />
image, and matrixes, that are the modern and conceptually simple copies<br />
of traditional photography.<br />
These two directions of research first gave rise to realizations of laboratories<br />
(e.g. HRSC sensor using several linear CCDs of the DLR, matrix<br />
cameras of the IGN-F), then to industrial realizations (ADS 40 of LH<br />
Systems in 2000 and MDC of Zeiss Intergraph in 2001). Studies have
Figure 1.5.1 (a) The digital modular camera DMC 2001 from Zeiss-Intergraph.<br />
Figure 1.5.1 (b) Three digital cameras from IGN-F: (from left to right) with one<br />
(1995), with two and with three separate cameras (2000), using<br />
KAF 6300 (2k × 3k pixels) or KAF 16800 (4k × 4k), or KAF<br />
16801 (4k × 4k with antiblooming) chips from Kodak.<br />
Figure 1.5.1 (c) The ADS-40 digital camera from Leica-Helava (left), the CCD<br />
lines observing various wavelengths (centre) and the optical<br />
subsystem for the separation of the visible light into 3 colours<br />
RGB (right).
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
shown quite clearly fields of interest of these two techniques, as well as<br />
their limitations. They have also shown how much the community of<br />
photogrammetrists had difficulties imagining the future with tools very<br />
different from older ones: this community is ready to replace one link of<br />
the photogrammetric chain at a time, but not much more, which is easily<br />
comprehensible. Photogrammetrists have, for example, clearly shown their<br />
concern, not because of directly digital images, but because of the entirely<br />
new situations where their empirical ability would have to be reconstituted<br />
from zero, in particular being entirely at the mercy of software<br />
editors. For example, linear CCD systems give a quite different geometry<br />
to process the images as compared to traditional techniques.<br />
The digital cameras developed according to these two general types are<br />
presented briefly below. They have in common the use of the CCD, which<br />
offers an extraordinary luminous response linearity compared with the<br />
usual argentic processes: these are really radiometers, perfectly calibratable,<br />
that make a much easier job, for example at the radiometric<br />
connections between neighbouring images at the time of the realization of<br />
mosaics or orthophotographies. The advantage of this linearity is also very<br />
clear when one has to suppress by digital processes the atmospheric<br />
diffusing fog: this operation is extremely simple and efficient with the CCD<br />
sensors, whereas it is impossible to automate it on argentic colour images.<br />
Currently the two solutions (linear CCD and matrixes) seem to have<br />
slightly different domains of applications, and as we already evoked it in<br />
§1.4, linear CCD systems are preferably destined to provide images of very<br />
high resolution but whose metrics is not very critical (for the merely visual<br />
exploitation and without precise measurements, for example, which represents<br />
a very large part of the market of aerial imagery). On the other hand<br />
the matrix systems, providing an extremely rigorous geometry (and even<br />
much better than that of argentic pictures, digitized or not), are far more<br />
suited to photogrammetric uses. But only after several years of market<br />
reaction to these offers can we make valid conclusions, and it is not possible<br />
for the authors to push the analysis farther at the time of writing this<br />
book. Some elements already presented in §1.4 will not be repeated here,<br />
and we refer the reader back to this section if necessary.<br />
1.5.2 CCD sensors in line<br />
<strong>Digital</strong> image acquisition 41<br />
It was in order to prepare a mission to Mars, that finally didn’t fly, that<br />
the DLR (Germany) studied in detail the HRSC tool (high resolution stereo<br />
camera). This tool, which was conceived for the spatial applications, was<br />
then tried with success on airborne missions. It has been used industrially<br />
since 1999 in production within the Istar company (France), and the new<br />
camera ADS 40 of LH Systems is based on the same concepts.<br />
The principle is as follows: in the focal plane of high-quality optics is<br />
installed a set of linear CCDs, which allows at any given instant the ground
42 Michel Kasser<br />
to be imaged in several directions. The prototype of the ADS 40 (Fricker,<br />
et al. 1999) uses four linear CCD or groups of linear CCD of 12,000<br />
elements each, one that gets images to the vertical of the plane, with the<br />
others imaging following a fixed angle with respect to the vertical (several<br />
angles are possible, according to the uses) toward the front and the rear<br />
of the plane. Thus, a point of the ground is seen successively under three<br />
different angles, which quite obviously permits one stereoscopic reconstitution<br />
of the relief. Some of these sensors work in panchromatic mode,<br />
others are equipped with spectral filters. Some linear CCD used in panchromatic<br />
mode are formed of two staggered linear CCD, separated by half a<br />
pixel, which may allow one to rebuild the equivalent of the signal provided<br />
by a linear CCD of 24,000 pixels. One then has a resolution that is without<br />
any comparison with any other airborne sensor used, even though the<br />
geometric quality of such images allows some minor distortions.<br />
Considering the trajectory of the plane, that is evidently uncertain and<br />
subject to unforeseeable and important movements, it is necessary to know<br />
at any instant and with a very high precision the real position of sensors in<br />
the space. This problem, which does not appear on a satellite (whose<br />
trajectory is extremely steady), implies for a plane the use of powerful additional<br />
systems capable of correcting at different times the effects of movements<br />
of the plane on the recorded image, that would be completely<br />
unusable without it. For that purpose, one attaches to the system an inertial<br />
platform as well as a GPS sensor; the fusion of the output data of these two<br />
sources is made exploiting a Kalman filter. Let us note that the coupling<br />
inertial platform GPS doesn’t necessarily permit one to determine in a<br />
satisfactory way the yaw movements of the plane: it would be determined<br />
correctly by the Kalman filter only if the plane did not remain for a long<br />
time right in line, which is unfortunately unavoidable. And long-distance<br />
systematisms of the GPS (for example, if it is used in trajectography mode),<br />
are directly transferred as deformations of the survey performed. Besides,<br />
one should not overestimate the operational performances of the GPS, and<br />
very short losses of signal can occur in flight, which leads at the time of the<br />
treatment to short interruptions of the corrections provided by the inertial<br />
platform: it almost produces automatically a set of irretrievable distortions<br />
of the image acquired. Therefore it is not foreseeable to get a geometric<br />
precision, expressed in pixels, comparable with what the digital images<br />
usually provide with matrix imagers in conical perspective. Typically one<br />
can count, in equivalent cases, on residual errors of several pixels for images<br />
obtained with linear CCD sensors, against 0.1 pixels on matrix images.<br />
The results published in 2000 show that the pros and cons of this sensor<br />
type ADS 40 include typically:<br />
• The possibility of an image with a number of distinct points on the<br />
ground close or even superior to what is available in aerial digitized<br />
pictures (12,000 pixels are roughly equivalent of what provides an
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
aerial picture 24 × 24 cm digitized with 20 m of pixel, and 24,000<br />
is achievable if necessary).<br />
• The geometric distortions due to inertial system insufficiencies are<br />
weak, but remain generally visible (some pixels): the metrics of the<br />
image therefore shows some small defects.<br />
• The dynamics of the image is satisfactory, but the necessity to transfer<br />
charges in the linear CCD to a very high speed implies a very limited<br />
exposure time, which does not allow dynamics as high as that on<br />
matrixes to be reached. Images, nevertheless, offer an excellent linearity<br />
of response.<br />
• The geometry of images (cylindro-conic perspective) implies new modes<br />
of digital data processes, which seem quite suited to the automated<br />
determination of digital surfaces models (DSM), but forbid the use of<br />
standard softwares of classic photogrammetry stations.<br />
• The restitution of the colour is obtained by adding in the focal plane<br />
additional linear CCDs equipped with filters. For example, the linear<br />
CCD that images along the vertical is in fact replaced by three parallel<br />
and very close linear CCDs equipped with the three usual filters, RGB.<br />
Other linear CCDs can be equipped with other filters, e.g. for the infrared.<br />
But if the three colours are obtained through lines that do not<br />
observe exactly the same area simultaneously, some artefacts appear on<br />
mobile objects, of course, but also when strong perspective differences<br />
are present in the image (high buildings, for example).<br />
1.5.3 Matrix CCD sensors<br />
<strong>Digital</strong> image acquisition 43<br />
The principle consists in preserving the usual geometry of the classic image<br />
acquisition (conical perspective), and using a CCD matrix as large as<br />
possible in the focal plane of the optics. Studies performed at IGN-F from<br />
the beginning of the 1990s (Thom and Souchon, 1999) show how such<br />
images could be inserted in photogrammetric classic chains, and what were<br />
the limitations and advantages of such matrix cameras:<br />
• The CCD matrixes are perfectly suitable for the compensation of linear<br />
blurring due to the speed (forward motion compensation or FMC), and<br />
even with more simplicity than on traditional FMC airborne cameras<br />
since it is performed electronically, without any mechanical movement.<br />
This allows a long exposure duration (several tens of milliseconds if<br />
need be), compatible with a very weak lighting (wintry or dusky image<br />
acquisitions) while preserving an excellent signal/noise ratio.<br />
• The sensitivity of the CCD, added to the possibility of using long exposure<br />
times, is the origin of a dynamics of the image that can reach<br />
considerable values, a digitization on 12 bits being hardly sufficient.<br />
It is quite different from silver halide film image performances that,<br />
even if digitized in the best conditions, would hardly deserve 6 bits.
44 Michel Kasser<br />
It is an essential asset, for example, for removing easily the atmospheric<br />
fog in post-process, or for observing in a satisfactory way shades<br />
and in the very luminous zones as well. It is also an essential asset for<br />
tools of automatic image matching, since with such a dynamic one no<br />
longer finds any uniform surface, and numerous details with quite faint<br />
differences of radiometry remain discernible.<br />
• The largest commercially available matrixes in 2000 have dimensions<br />
of the order of 7,000 × 9,000 pixels. This represents a more modest<br />
pixel number than that of a digitized silver halide image. Nevertheless,<br />
cases of the use of such images of very high dynamics have been valued<br />
on applications of current middle-scale cartography: an equivalent<br />
service to a given ground size of pixel on a digitized argentic image<br />
is more or less provided by a two times larger pixel on a digital image<br />
of very large dynamics. Compared to the service provided by a traditional<br />
digitized image, it corresponds, therefore, to a digital image<br />
about half as large. Let us note that with the remarkable performances<br />
of these images it would be quite acceptable and logical to<br />
interpolate pixels of dimension half as large, which would so give<br />
the equivalent of an image 8,000 × 8,000 from a matrix 4,000 ×<br />
4,000. One would thus get the content of an image quite similar to a<br />
digitized argentic one, although of incomparably better linear radiometry.<br />
Nevertheless, for a given site of study this may imply, according<br />
to the chosen matrix size, more axes of flight than for a classic image<br />
acquisition, which is a source of additional costs. Developments of<br />
large dimension matrixes, by chance, follow the considerable demand<br />
for digital pictures for the general public, so that they evolve rapidly.<br />
Apparently formats used by amateurs (24 × 36 mm) and professionals<br />
(60 × 60 mm) are currently well covered by matrixes of 2,000 × 3,000<br />
and 4,000 × 4,000. The likely evolutions could go circa a strong reduction<br />
of costs, the trichromy, and maybe an increase of current matrix<br />
sizes, but not necessarily very important (rather for the professional<br />
photographers, who represent a small section of the market) knowing<br />
that the data storage is no longer a limitation for these markets due<br />
to the availability of large data storage capabilities. The way chosen<br />
by the ZI society (DMC camera) consists in a combination of various<br />
modular matrix subsets based on matrixes about 4,000 × 7,000,<br />
capable of reconstituting 8,000 × 14,000 images, with the possibility<br />
of supplementary matrixes equipped with filters capable of reconstituting<br />
of colours. The elementary matrixes are equipped with individual<br />
optics aiming to divergent axes and assuring a small overlap between<br />
each of the four images. The resulting geometry is equivalent, within<br />
an accuracy of about 0.1 pixels, to an image acquired in only one<br />
block, by one optic.<br />
• The restitution of the colour can be obtained of two ways: either on<br />
certain matrixes a set of filters is deposited on the pixels, or one uses
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
<strong>Digital</strong> image acquisition 45<br />
three matrixes equipped separately with filters. The first solution<br />
induces delicate problems of reconstitution of the colour, and requires<br />
well-developed software to let only small artefacts (e.g. on borders, or<br />
on very local details displayed on only one pixel). It implies either a<br />
subset of 4 pixels (e.g. Kodak: two with green, one with blue and one<br />
with red filters), or a different arrangement of the matrix, where the<br />
lines are staggered from a half-pixel, and the 3 filters are regularly<br />
spaced on a 3-pixel subset (e.g. Fuji: on each line the filters are R G<br />
B R G B . . ., for example the R being staggered by 1.5 pixels from<br />
a line to the other), a solution that provides a better colour synthesis.<br />
The second solution is costlier, because it finally results in the equivalent<br />
of three complete cameras regrouped together, but it permits the<br />
perfect reconstitution of colours in all circumstances.<br />
• The metrics of the image is absolutely excellent, and even significantly<br />
superior to the accessible metrics on classic film cameras. Even if the<br />
latter are nearly perfect in terms of distortion, the film represents a<br />
medium whose distortions are far from negligible. These distortions<br />
are due to treatments at the time of the development and the drying,<br />
that stretch the film in an anisotropic way. They are also, and this<br />
phenomenon is not so well known (see §1.1), due to dust on the<br />
bottom plate of the camera that prevents a good contact of the film<br />
on the table to depression, and creates some significant bumps at the<br />
time of the image acquisition. It provokes local distortions of the image<br />
that are pure artefacts, nearly impossible to model. In the digital matrix<br />
cameras, the calibration is performed as for a classic camera, but as<br />
the sensor is monolithic, no distortion can now occur, and one observes<br />
a precision in geometric computations that is generally better that 0.1<br />
pixels.<br />
• The available optics offers a much wider choice for amateur cameras<br />
than for the classic aerial cameras. Practically all optics adapted to the<br />
60 × 60 format (professional photography) are satisfactory, and this<br />
at obviously much more modest costs. Indeed the classic defects of<br />
optics, distortion, vignetting, etc., can be corrected by a posteriori<br />
calculation, and only the resolution of the optics and its dimensional<br />
stability in the time remain important parameters.<br />
1.5.4 Specifications of digital aerial image acquisitions<br />
The use of airborne digital cameras induces significant changes in terms<br />
of specification of image acquisitions. Notions of height of flight and<br />
size of pixel can be considered as largely independent, in particular due<br />
to the much larger choice of available optics. Even the notion of scale<br />
disappears completely, with all the empirical knowledge that is attached<br />
to it. The maximal angle of field of the image will be chosen according<br />
to the acceptable hidden parts and the need of precision on an altimetry
46 Michel Kasser<br />
restitution, expressed in particular through the B/H ratio of the distance<br />
between successive positions of the camera at the time of two successive<br />
image acquisitions, divided by the altitude of flight. The size of pixel will<br />
be chosen according to the size and the nature of objects to survey and<br />
to detect, and to the dynamics of the accessible image with the considered<br />
camera. It is with these two parameters that one will determine the project<br />
of flight, and in particular the altitude.<br />
One will note some empirical rules that will be detailed in §1.9: the precision<br />
of pointing of an object is performed in general to better than 0.3<br />
pixels, and to identify an object easily it is desirable that it is not smaller<br />
than 3 × 3 pixels. This allows one to specify the really necessary pixel sizes.<br />
For example, with a matrix camera with a very large dynamics, to achieve<br />
an orthophotography on an urban zone where a DTM is already available,<br />
one will choose a long focal distance to limit the hidden parts of buildings,<br />
and the raw size of pixel will be chosen, e.g. as 50 cm to have a satisfactory<br />
description of buildings, or of 30 cm if one wants to have a tool<br />
of management of urban furniture. These sizes of pixels will be reduced if<br />
the dynamics of the image is low (respectively 30 cm and 15 cm for a digitized<br />
classical image).<br />
In a traditional image acquisition the height of flight also intervenes by<br />
the atmospheric fog, more or less acceptable, that will result: sometimes,<br />
to get a very weak fog with a height superior to 4,000 m may prove to<br />
be an impossible mission under certain latitudes in classic photography.<br />
In digital imagery, if the dynamics is sufficient it allows the removal of<br />
the atmospheric fog while preserving a good quality of colour and a good<br />
capacity to work in zones of shades. It permits a considerable reduction<br />
in the delays bound to the meteorological risks, but also eases the constraint<br />
on the parameter of flight height, if need be.<br />
References<br />
Fricker P., Sandau R., Walker S. (1999) <strong>Digital</strong> Aerial Sensors: possibilities and<br />
problems, Proceedings of OEEPE Workshop on Automation in <strong>Digital</strong><br />
Photogrammetric Production, Marne-la-Vallée, 22–24 June, Public. OEEPE no.<br />
37, pp. 81–90.<br />
Thom C., Souchon J.-P. (1999) The IGN <strong>Digital</strong> Camera System, Proceedings of<br />
OEEPE Workshop on Automation in <strong>Digital</strong> Photogrammetric Production,<br />
Marne-La-Vallée, 22–24 June, Public. OEEPE no. 37, pp. 91–96.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
1.6 RADAR IMAGES IN PHOTOGRAMMETRY<br />
Laurent Polidori<br />
1.6.1 Overview<br />
Images obtained from a synthetic aperture radar (SAR) are very sensitive<br />
to terrain undulations, and this sensitivity has been the basis for the development<br />
of several relief-mapping techniques using one or several radar<br />
images, namely, radargrammetry, interferometry and radarclinometry<br />
(Polidori, 1991). Each of these techniques was initially proposed more than<br />
25 years ago, and their theoretical bases were established in the 1980s. The<br />
state-of-the-art of radar mapping has been presented in detail by Leberl<br />
(Leberl, 1990), based on airborne campaigns and on pioneering spaceborne<br />
SARs like SEASAT (1978), SIR-A (1981) and SIR-B (1984), but the actual<br />
potential of radar techniques has been known only for a few years, due to<br />
the lack of SAR data before the 1990s. The launch of several spaceborne<br />
SARs in recent years (ERS, JER-S, RADARSAT, SIR-C) has led to further<br />
experiments and to a better estimation of the accuracy of radar-derived<br />
DSMs (digital surface models).<br />
Several specific properties of radar sensors must be mentioned to explain<br />
their potential for mapping.<br />
An active device<br />
A radar sensor in general, and a synthetic aperture radar in particular,<br />
provides its own power through a transmitting antenna: this contributes<br />
to making radar a so-called ‘all-weather’ sensor by allowing acquisitions<br />
even at night. Moreover, since day and night acquisitions correspond to<br />
ascending and descending orbits, this also contributes to opposite side<br />
viewing capabilities so that hidden areas are greatly reduced.<br />
Absolute location accuracy<br />
Radar images in photogrammetry 47<br />
Attitude (i.e. roll, yaw and pitch angles), which in the case of optical<br />
images is the major contributor of absolute location error, has no effect<br />
on location in a SAR image. Indeed, the computation of image point location<br />
in a SAR image requires only a slant range and a Doppler frequency<br />
shift, and these magnitudes do not depend on sensor attitude.<br />
Sensitivity to relief<br />
Radar image location is based on slant range measurements between<br />
the antenna and each terrain target: this is why terrain elevation has a<br />
direct effect on image location. This effect can be locally described as
48 Laurent Polidori<br />
a proportional relationship between a height variation z and a slant range<br />
variation R for a given incidence angle :<br />
R z<br />
cos ,<br />
where is the angle between the local vertical and the viewing direction.<br />
At pixel scale, R is a parallax that can be measured to derive an<br />
elevation: this is the basis for radargrammetry. At wavelength scale, R<br />
can be estimated from a phase shift between two echoes: this is the basis<br />
for interferometry. Apart from this geometrical sensitivity, a radiometric<br />
sensitivity can be mentioned, since terrain orientation has an impact on<br />
radar return intensity: this is the basis for radarclinometry (or radar shape<br />
from shading).<br />
Coherence<br />
Radar signals are coherent, which means that they are determined in terms<br />
of amplitude and phase. The generation of a SAR image would not be<br />
possible otherwise. This characteristic makes radar sensors very sensitive<br />
to relief because of interferometric capabilities.<br />
Atmospheric effects<br />
The effects of the atmosphere and in particular the troposphere can be<br />
neglected at pixel scale, i.e. a radar image can be acquired and accurately<br />
located whatever the meteorological conditions. On the contrary, these<br />
effects cannot be neglected at wavelength scale, and they produce artefacts<br />
in interferometric products. For instance, a radar echo acquired under<br />
heavy rain conditions may have a phase delay of several times 2 even if<br />
the scene geometry has not changed.<br />
1.6.2 Radargrammetry<br />
Radargrammetry is an adaptation of the photogrammetric principles to<br />
the case of radar images. Indeed, it is based on parallax measurements<br />
between two images acquired from different viewpoints. However, radargrammetry<br />
cannot be carried out directly with photogrammetric tools, for<br />
two main reasons.<br />
First, the SAR geometry is modelled by specific equations: this implies<br />
that analogue stereo plotters are not suitable, and that the first rigorous<br />
implementations could only be achieved with analytical plotters (Raggam<br />
and Leberl, 1984).<br />
The second reason is that stereo viewing is difficult and uncomfortable<br />
with radar images, in particular in the case of rugged terrain. This is due
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
to the fact that bright points and shadows suffer migration from one image<br />
to another due to the difference in illumination. However, stereo viewing<br />
capabilities can be acquired with practice, as relief perception is basically<br />
a psychological process (Toutin, 1997a).<br />
The feasibility of radargrammetry was demonstrated for airborne radar<br />
data (Leberl et al., 1987) as well as for spaceborne radar data (Leberl<br />
et al., 1986). According to experiments made over different landscapes, it<br />
is now well known that this technique can provide 3D measurement with<br />
an accuracy of a few pixels, i.e. around 20 m with airborne SAR (Toutin,<br />
1997b) and between 20 and 50 m with ERS (Toutin, 1996; Raggam et al.,<br />
1993) or RADARSAT (Sylvander et al., 1997; Marinelli et al., 1997). As<br />
expected, the radargrammetric error tends to increase in case of steep relief.<br />
1.6.3 Interferometry<br />
Radar images in photogrammetry 49<br />
SAR interferometry consists in deriving terrain elevation from the phase<br />
difference between two radar echoes acquired from very close antenna<br />
positions (Zebker and Goldstein, 1986; Massonnet and Rabaute, 1993).<br />
The two images can be obtained either from two antennas constrained to<br />
remain on parallel paths (this is generally the case for airborne systems)<br />
or from the same antenna during two passes. If these images are properly<br />
registered, the phase difference can be computed for every pixel and stored<br />
in an interferogram. Although the phase difference has a simple relationship<br />
with the slant range difference and therefore with elevation, relief<br />
cannot be derived so easily due to two severe limitations.<br />
The first limitation is the dependence of radar phase on non-topographic<br />
factors, such as atmospheric variations (Goldstein, 1995; Kenyi and<br />
Raggam, 1996), land cover changes or sensor miscalibration (Massonnet<br />
and Vadon, 1995): these contributions to phase shift contaminate the interferogram<br />
and can be converted to elevation, contributing to the error of<br />
the output digital surface model. This is generally the case when the time<br />
interval between the acquisitions is too long, so that important changes<br />
have occured not only in the atmospheric refraction index but also in the<br />
surface roughness or electromagnetic properties.<br />
The second limitation is the fact that the radar phase is not known in<br />
absolute terms but only modulo 2, so that interferograms generally exhibit<br />
fringes that must be unwrapped.<br />
The performance of phase unwrapping mainly depends of the B/H ratio,<br />
i.e. the ratio between the stereoscopic baseline and the flight height:<br />
• when B/H is smaller, the fringes are wider: it becomes easier to unwrap<br />
them, but a given height gradient has less impact on the radar phase,<br />
which means that the interferogram is less sensitive to relief;<br />
• when B/H is larger, the fringes become narrow and noisy: they are<br />
too sensitive to relief and they are more difficult to unwrap.
50 Laurent Polidori<br />
In the most usual configuration, interferometric images are acquired with<br />
a single antenna at different dates. The problem then is that the baseline<br />
cannot be predicted. Therefore, the accuracy of the method, which can be<br />
determined a posteriori, cannot be predicted before the acquisitions.<br />
The most accurate results, which are close to the theoretical limits, are<br />
generally obtained with simultaneous dual antenna acquisitions. Interferometric<br />
processing based on the airborne TOPSAR system (Zebker et al.,<br />
1992) provided accuracies around 1 m in flat areas and 3 m in hilly<br />
areas (Madsen et al., 1995), and the CCRS dual-antenna SAR system<br />
provided a 10 m accuracy in mountainous and glacial areas (Mattar<br />
et al., 1994).<br />
Single-antenna interferometry provides very heterogeneous results, due<br />
to the double influence of the spatial baseline (which cannot be predicted)<br />
and the time interval (which is generally too long) (Vachon et al., 1995).<br />
These influences vary according to landscape characteristics, and the accuracy<br />
depends on slope and land cover. Since these characteristics are often<br />
heterogeneous over 100 km side scenes, evaluating an interferometric accuracy<br />
over such a huge area is not very meaningful. Under suitable conditions<br />
(moderate relief, short time between acquisitions, no atmospheric variations),<br />
the error of interferometric products may be as small as 5 m with<br />
ERS data (Dupont et al., 1997) and 7 m with RADARSAT data (Mattar<br />
et al., 1998). As soon as ideal conditions are not fulfilled, errors often<br />
increase to several tens of metres.<br />
During the generation of a radar interferogram, an associated product<br />
called the coherence image is usually computed as well, in order to display<br />
the correlation between the two complex echoes. This product provides<br />
useful information on the interferometric accuracy, because phase noise<br />
increases the output height error proportionally. However, correlation is<br />
also a very interesting thematic product, since correlation is an indicator<br />
of surface stability. For instance, rocky landscape and urban structures are<br />
very stable, so that they are generally mapped with high interferometric<br />
correlation, while forested areas and, above all, water, have very low correlation<br />
because the surface geometry within each pixel has permanent<br />
changes comparable to the wavelength (Zebker and Villasenor, 1992).<br />
1.6.4 Radarclinometry (shape from shading)<br />
While radargrammetry and interferometry are based on the principle of<br />
stereoscopy, and therefore on the geometry of the radar images, radarclinometry<br />
makes use of image intensity and can provide a DSM using a<br />
single radar image.<br />
Radarclinometry is a particular case of shape from shading, insofar as<br />
it determines the absolute orientation of each terrain facet using its intensity<br />
in the radar image. In fact, this technique is a quantitative application<br />
of the visual perception of relief one has looking at a radar image.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
The link between terrain orientation and radar image intensity is<br />
described by a backscattering model, in which the intensity is expressed<br />
as a function of the incidence angle (between the viewing direction and<br />
the perpendicular to the terrain surface) and the backscattering coefficient<br />
(similar to reflectance for optical imagery). Radarclinometry is basically<br />
an inversion of such a model. However, inverting a backscattering model<br />
is constrained by two major limitations:<br />
• image intensity depends not only on the incidence angle but also on the<br />
backscattering coefficient and therefore on ground characteristics (land<br />
cover, roughness, moisture . . .) which in most cases are not known;<br />
• one particular value of the incidence angle corresponds to an infinity<br />
of possible terrain orientations, because surface azimuth or aspect also<br />
contributes to the incidence angle.<br />
Due to these ambiguities, any radarclinometric algorithm must inject<br />
external constraints or hypotheses in order to make the inversion possible.<br />
To mention the most classical hypotheses, the first ambiguity is generally<br />
solved by assuming a homogeneous land cover (which implies that intensity<br />
variations are caused by slope variations only) and the second one by<br />
neglecting the slope component oriented along the track. A number of<br />
algorithms have been proposed to overcome these difficulties (Frankot<br />
and Chellapa, 1987; Thomas et al., 1989; Guindon, 1990; Paquerault and<br />
Maître, 1997).<br />
Guindon (1990) obtained an elevation above 200 m from SEASAT data<br />
(with around 20°, i.e. a very deep incidence) over a mountainous area<br />
with steep slope: these are the worst conditions for radarclinometry. On<br />
the contrary, Paquerault and Maître (1997) obtained an error of the order<br />
of 20 m using RADARSAT over French Guyana, where the landscape is<br />
uniformly forested with gentle hills: this corresponds to ideal conditions<br />
for radarclinometry.<br />
Finally, it should be noted that the elevation error is not very meaningful<br />
to evaluate the performance of radarclinometry, which is bascally<br />
a slope-mapping technique. The fact that standard elevation error is often<br />
used as a unique quality criterion for DSMs implies that very smooth<br />
surfaces are generally preferred, while radarclinometry is very sensitive to<br />
microrelief.<br />
References<br />
Radar images in photogrammetry 51<br />
Dupont S., Nonin P., Renouard L. (1997) Production de MNT par interférométrie<br />
et radargrammétrie. Bull. SFPT, no. 148, pp. 97–104.<br />
Frankot R., Chellapa R. (1987) Application of a shape-from-shading technique to<br />
synthetic aperture radar. Proceedings IGARSS ’97 (Ann Arbor), pp. 1323–<br />
1329.
52 Laurent Polidori<br />
Goldstein R. (1995) Atmospheric limitations to repeat-pass interferometry. Geophysical<br />
Research Letters, vol. 22, no. 18, pp. 2517–2520.<br />
Guindon B. (1990) Development of a shape-from-shading technique for the extraction<br />
of topographic models from individual spaceborne SAR images. IEEE<br />
Transaction on Geoscience and Remote Sensing, vol. 28, no. 4, pp. 654–661.<br />
Kenyi L., Raggam H. (1996) Atmospheric induced errors in interferometric DEM<br />
generation. Proceedings of IGARSS ’96 Symposium (Lincoln), pp. 353–355.<br />
Leberl F. (1990) Radargrammetric image processing. Artech House, Norwood,<br />
p. 613.<br />
Leberl F., Domik G., Raggam H., Kobrick M. (1986) Radar stereomapping techniques<br />
and application to SIR-B images of Mt. Shasta. IEEE Transaction on<br />
Geoscience and Remote Sensing, vol. GE-24, no. 4, pp. 473–481.<br />
Leberl F., Domik G., Mercer B. (1987) Methods and accuracy of operational digital<br />
image mapping with aircraft SAR. Proceedings of the Annual Convention of<br />
ASPRS, vol. 4, pp. 148–158.<br />
Madsen S., Martin J., Zebker H. (1995) Analysis and evaluation of the NASA/<br />
JPL TOPSAR across-track interferometric SAR system. IEEE Transactions on<br />
Geoscience and Remote Sensing, vol. 33, no. 2, pp. 383–391.<br />
Marinelli L., Toutin T., Dowman I. (1997) Génération de MNT par radargrammétrie.<br />
Bull. SFPT, no. 148, pp. 89–96.<br />
Massonnet D., Rabaute T. (1993) Radar interferometry: limits and potential. IEEE<br />
Transaction on Geoscience and Remote Sensing, vol. 31, no. 2, pp. 455–464.<br />
Massonnet D., Vadon H. (1995) ERS-1 internal clock drift measured by interferometry.<br />
IEEE Transaction on Geoscience and Remote Sensing, vol. 33, no. 2,<br />
pp. 401–408.<br />
Mattar K., Gray L., Van der Kooij M., Farris-Manning P. (1994) Airborne interferometric<br />
SAR results from mountainous and glacial terrain. Proceedings<br />
IGARSS’94 Symposium (Pasadena), pp. 2388–2390.<br />
Mattar K., Gray L., Geudtner D., Vachon P. (1998) Interferometry for DEM and<br />
terrain displacement: effects of inhomogeneous propagation Canadian Journal<br />
of Remote Sensing, vol. 25, no. 1, pp. 60–69.<br />
Paquerault S., Maître H. (1997) La radarclinométrie. Bull. SFPT, no. 148, pp. 20–29.<br />
Polidori L. (1991) <strong>Digital</strong> terrain models from radar images: a review. Proceedings<br />
of the International Symposium on Radars and Lidars in Earth and Planetary<br />
Sciences (Cannes), pp. 141–146.<br />
Raggam H., Leberl F. (1984) SMART – a program for radar stereo mapping on<br />
the Kern DSR-1. Proceedings of the Annual Convention of ASPRS, pp. 765–773.<br />
Raggam H., Almer A., Hummelbrunner W., Strobl D. (1993) Investigation of the<br />
stereoscopic potential of ERS-1 SAR data. Proceedings of the 4th International<br />
Workshop on Image Rectification of Spaceborne Synthetic Aperture Radar<br />
(Loipersdorf, May) pp. 81–87.<br />
Sylvander S., Cousson D., Gigord P. (1997) Etude des performances géométriques<br />
des images RADARSAT. Bull. SFPT, no. 148, pp. 57–65.<br />
Thomas J., Kober W., Leberl F. (1989) Multiple-image SAR shape from shading.<br />
Proceedings IGARSS ’89 (Vancouver, July), pp. 592–596.<br />
Toutin Th. (1997a) Depth perception with remote sensing data. Proceedings of<br />
the 17th Earsel Symposium on Future Trends in Remote Sensing (Lyngby, June)<br />
pp. 401–409.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Toutin Th. (1997b) Accuracy assessment of stereo-extracted data from airborne<br />
SAR images. International Journal for Remote Sensing, vol. 18, no. 18, pp.<br />
3693–3707.<br />
Vachon P., Geudtner D., Gray L., Touzi R. (1995) ERS-1 synthetic aperture radar<br />
repeat-pass interferometry studies: implications for RADARSAT. Journal Canadien<br />
de Télédétection, vol. 21, no. 4, pp. 441–454.<br />
Zebker H., Goldstein R. (1986) Topographic mapping from interferometric SAR<br />
observations. Journal of Geophysical Research, vol. 91, no. B5, pp. 4993–4999.<br />
Zebker H., Villasenor J. (1992) Decorrelation in interferometric radar echoes. IEEE<br />
Transaction on Geoscience and Remote Sensing, vol. 30, no. 5, pp. 950–959.<br />
Zebker H., Madsen S., Martin J., Wheeler K., Miller T., Lou Y., Alberti G.,<br />
Vetrella S., Cucci A. (1992) The TOPSAR interferometric radar topographic<br />
mapping instrument. IEEE Transaction on Geoscience and Remote Sensing,<br />
vol. 30, no. 5, pp. 933–940.<br />
1.7 USE OF AIRBORNE LASER RANGING SYSTEMS FOR<br />
THE DETERMINATION OF DSM<br />
Michel Kasser<br />
Airborne laser ranging systems for DSM 53<br />
1.7.1 Technologies used, performances<br />
Airborne laser telemetry (ALRS, an acronym that stands for airborne laser<br />
ranging system) is the normal term used in the 1970s for the airborne<br />
profile recorders (APR) that provided, in a continuous way, a very precise<br />
vertical distance from the airplane to the ground (for example Geodolite®<br />
of Spectra-Physics); the positioning of the plane was obtained then by<br />
photographic means that were not very precise in planimetry, the altimetric<br />
reference being directly the isobar surface at the level of the plane.<br />
These processes, fallen into obsolescence since the beginning of the 1980s<br />
because of insufficient precision of positioning, regained an interest when<br />
the GPS appeared as a very precise localization system, with data being<br />
processed after the flight. A large variety of equipment is available in 2000,<br />
including the laser rangefinder, its scanning device, an inertial platform<br />
and a GPS receiver, and of course the originality of every equipment resides<br />
in the proposed software packages.<br />
The principle of functioning is as follows:<br />
• The ALRS provides in a continuous way, at rates from 2 to 100 kHz,<br />
the time of propagation of a laser impulse given out by a powerful<br />
laser diode working in the near IR. This laser rangefinder is fixed to<br />
the plane as rigidly as possible. It may, in general, according to requirements,<br />
provide some different distances, for example that on the first
54 Michel Kasser<br />
received echo (to measure objects over the ground, like an electric<br />
power line), or on the contrary on the last (to measure ground under<br />
vegetation), or again all echoes received for each shot (but then the<br />
quantity of data greatly increases). The precision of the rangefinder is<br />
in the order of some centimetres in the absolute, but in fact what<br />
matters is the interaction of the laser beam with the object targets,<br />
which is geometrically not very definite in most cases (e.g. high vegetation,<br />
building borders . . .). The beam is slightly diverging (typically<br />
1 mrad), and even though the plane cannot fly high, considering the<br />
weak optic signal that is allowed not to create any ocular hazards, the<br />
analysis spot therefore has a diameter of several decimetres. A variant<br />
of this technology must be mentioned; it uses a continuous laser source<br />
modulated by a set of frequencies permitting ambiguity resolution (on<br />
the model of the old Geodolite® already quoted), but it is not widely<br />
used.<br />
• An optico-mechanical scanner device allows the laser beam to be sent<br />
sequentially in different directions, which permits, with the advancement<br />
of the plane a scan of the ground, in a strip around the ground<br />
trace to be performed.<br />
• In order to know the position of every laser shot in the space in spite<br />
of the unknown movements of the plane, the ALRS is coupled rigidly<br />
to an inertial unit platform. It provides information at a rate compatible<br />
with the movements of the plane, typically from 200 to 500 Hz.<br />
It must be initialized (some minutes or tens of minutes), so that its<br />
core components are in a stationary thermal regime. Its data errors<br />
are characterized by a bias proportional to the square of the time,<br />
which implies very frequent corrections.<br />
• In order to correct the inertial unit that would drift quickly to excessive<br />
values without such external help, a precise GPS receiver is added<br />
to the whole. This receptor provides, in differential mode, the positions<br />
of its antenna as often as possible, generally between rates of 1 and<br />
10 Hz, and the post-processing must provide precision compatible with<br />
the requirements of the study, which in turn implies naturally the localization<br />
of the reference GPS receiver in relation to the surveyed area.<br />
• To complement this equipment, there is a PC unit to pre-process data<br />
and store the large quantities of observed data, to help with the navigation,<br />
etc.<br />
The post-processing of measures is an intense operation. The GPS calculation,<br />
then the fusion of the GPS with inertial data, is an operation that<br />
may be automated to a large extent. Then it is necessary to process echoes,<br />
that is to say to filter non-applicable echoes, and this operation is not<br />
currently automated at all. For example to get a DTM under trees, it is<br />
necessary to identify ‘by hand’ which echoes probably come from the
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
ground and which do not. Considering the number of echoes to treat,<br />
this operation, even with the help of the software, is very tedious. The<br />
impersonal and systematic character of the target selection (whatever the<br />
target, there is always an echo) is one of the fundamental features of<br />
this survey mode, and it may imply heavy data post-processing in certain<br />
configurations.<br />
1.7.2 Comparisons with photogrammetry<br />
Airborne laser ranging systems for DSM 55<br />
Surveys by ALRS are significantly different from photogrammetric surveys,<br />
either using traditional digitized images or direct digital imagery.<br />
These surveys are in a semi-experimental period, in which the distributed<br />
and published information are not necessarily all applicable considering<br />
the commercial implications. ALRS surveys have therefore a<br />
low maturity compared to photogrammetry. One is nevertheless able to<br />
identify common features and weaknesses of these two techniques. An<br />
excellent survey on this topic is that by Baltsavias (1999), from which<br />
we take here certain elements. The main comparison between ALRS and<br />
photogrammetric surveys must include types of data process that are at a<br />
semi-experimental level in the two domains, which includes therefore, for<br />
example, developments in progress concerning automatic image matching<br />
and multi-stereoscopy.<br />
Points that we will keep in this comparison are as follows:<br />
• The DTM issued from ALRS can be available, if the type of work is<br />
suitable, in a very short time after the flight. For example, for the<br />
linear sites surveys without vegetation (urban surveys, power lines),<br />
the post process can be entirely automated, and the geometric data<br />
supplied in one or two hours. To the contrary, the supplying of DTM<br />
by automatic image matching requires processes that are not yet<br />
completely automatic, and are always quite long.<br />
• The equipment used in ALRS is sophisticated and expensive, and some<br />
unit components may experience some export limitations according to<br />
their countries of origin (USA for example). The maintenance when<br />
in use is therefore still not very easy. We are in the presence of technologies<br />
of which some are very recent, evidently in contrast with<br />
photogrammetry where devices of aerial image acquisition are industrialized<br />
at a very high level, and are of a remarkable reliability.<br />
• The installation of the ALRS equipment on board a photogrammetric<br />
plane or a helicopter is a delicate phase that requires, for example,<br />
aiming toward the ground, which is often not an easy situation. Besides,<br />
the ALRS equipment requires a GPS antenna (delicate to install on<br />
certain helicopters) and consumes electric power that is not always<br />
available on the current planes. The geometric link between different
56 Michel Kasser<br />
subsets (GPS antenna, inertial power station, ALRS) is a delicate<br />
metrology operation, but fortunately it can be checked by the airborne<br />
measures themselves if the software of calculation foresees it.<br />
• The survey by ALRS is an active technology that doesn’t require, in<br />
any aspect, lighting by the sun. There are no limitations therefore to<br />
timetables of flight, other than those dictated by the security of the<br />
aerial navigation. Flights are necessarily performed at low altitude,<br />
therefore in zones with little problem of flight authorization (out of<br />
the commercial air space), but which on the contrary are more and<br />
more difficult to get over cities. But this possibility of working in the<br />
morning, in the evening or in winter doesn’t present the same comparative<br />
advantage now that one may use digital cameras, whose light<br />
sensitivity is such that one can acquire excellent pictures without difficulty<br />
with a low level of ambient light (low sun, flight under clouds).<br />
As long as the question is to provide some DTM, such flights under<br />
weak lighting are quite acceptable and competitive with the ALRS,<br />
but the current problem is that one very often asks also for an aerial<br />
image acquisition to serve as a basis for photo-interpretation, and even<br />
the production of orthophotographies, and this requires good sun<br />
lighting on the other hand. However, the ALRS generally cannot<br />
provide a picture, in complement of the DTM: it is a tool to provide<br />
DTM and nothing else. It doesn’t allow one to discern in a simple<br />
way a house from a tree, for example, which limits its field of application<br />
quite a lot. And if one adds a camera in the plane, considering<br />
the height of flight it leads to a very significant image quantity, so that<br />
one would not know how to deal with the present means in order to<br />
make a mosaic or an orthophotography, even though all necessary<br />
data are available.<br />
• Interactions of the ALRS with vegetation are interesting and numerous.<br />
To the list of problems, let us note that echoes obtained in the<br />
high vegetation are very difficult to assign to a specific layer, that<br />
is one works with the first or the last echo detected. On the other<br />
hand, in forests the ALRS is the only technique that can provide,<br />
in practically all types of forests (even tropical), a DTM of the<br />
ground through leaves and branches. It presents a major interest<br />
therefore for planning in equatorial zones, where the hydrology is<br />
dominant and cannot be correctly processed without such data.<br />
(See Figure 1.7.1.)<br />
• An interesting feature of the ALRS resides in its capacity to measure<br />
some echoes even on very small objects. The most spectacular example<br />
is that of high-tension power line surveys. It would be impossible in<br />
photogrammetry to survey lines, not because one doesn’t see them<br />
(with a digital camera one sees easily even thin wires thanks to the<br />
considerable dynamics of the picture), but because there is no means<br />
to aim at the cable when the stereoscopic basis is parallel to the line,
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Airborne laser ranging systems for DSM 57<br />
Figure 1.7.1 Two raw profiles obtained by the ALTOA® ALRS in French<br />
Guyana (tropical vegetation around 40 m high). One notes that the<br />
shots ranging to the ground describe very finely the topography, and<br />
that in such circumstances a photogrammetric survey using the tops<br />
of the canopy and subtracting a constant value would (1) ignore<br />
important morphologic features such as trenches and (2) induce<br />
large errors as the height of the canopy is far from constant.<br />
which is necessarily the case for a linear survey of this type. In hightension<br />
power line surveys by ALRS, it is possible to survey almost in<br />
an entirely automated way lines and the zones of land and vegetation<br />
that are close, which permits the identification on time of the interventions<br />
of lopping that could be necessary.<br />
• In most other domains, there are few differences between obtaining a<br />
DTM by ALRS or by photogrammetry. Some authors consider that<br />
in urban situations the ALRS presents significant advantages, but one<br />
can note many counter-examples, and one may specify that the multiimage<br />
matching (automatic correlation with more than two images)<br />
on images of large dynamics in the city also provides excellent results,<br />
with advantages and drawbacks that are different.<br />
• We must note that error models are not at all the same, and that<br />
neither of the two techniques really surpasses the other. For example<br />
punctual mistakes are rare in photogrammetry, and numerous in ALRS<br />
(electric lines, birds, reflection on planes of water, etc.). The precision<br />
of measurement of a building may be also analysed in different ways<br />
in the two cases: the automatic matching will easily find the edges but<br />
will be in difficulty because of the hidden parts, the ALRS won’t<br />
describe the edges correctly because of the width of the analysis spot.
58 Michel Kasser<br />
For example, with regard to measures of pure altimetry, one cannot<br />
establish a net benefit in terms of precision, the ALRS being locally<br />
exact, but not being capable of identifying ditches or very important,<br />
although small, streams on such sites. The data provided by ALRS or<br />
by automatic image matching in photogrammetry are not originally<br />
structured. Numerous studies are in progress to structure them in an<br />
automatic way at least for the simple cases (e.g. buildings). And in<br />
this domain photogrammetric methods use extensively the possibilities<br />
of automatic analysis of every image, whereas the ALRS data are<br />
much more difficult to analyse.<br />
1.7.3 Perspectives, conclusions<br />
One is therefore in the presence of a new technology, whose optimal use<br />
in cases probably must be deepened again, but that one has nevertheless<br />
to take into account from now on for data production (Ackermann, 1999).<br />
The ALRS must be considered as a complement to photogrammetry, especially<br />
suitable in some cases:<br />
• surveys under vegetation;<br />
• linear sites DTM and very fast supplying of narrow zones;<br />
• high-voltage power line surveys;<br />
• bare surface DTM.<br />
There are a great many studies in progress to improve the fruitfulness of<br />
this technique, and costs will probably soon become clearer, considering<br />
their present instability.<br />
References<br />
Ackerman F. (1999) Airborne laser scanning – present status and future expectations.<br />
ISPRS Journal of <strong>Photogrammetry</strong> and Remote Sensing 54, pp. 64–67.<br />
Baltsavias P.E. (1999) A comparison between photogrammetry and laser scanning.<br />
ISPRS Journal of <strong>Photogrammetry</strong> and Remote Sensing 54, pp. 83–94.<br />
1.8 USE OF SCANNERS FOR THE DIGITIZATION OF<br />
AERIAL PICTURES<br />
Michel Kasser<br />
1.8.1 Introduction<br />
The present need for digital pictures destined for photogrammetric<br />
processing is probably not going to stop growing in the next few years.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Scanners for aerial image digitization 59<br />
It was at first bound to developments in the digital orthophotography<br />
market, from the 1980s. This development is far from finished, especially<br />
because users are now able to handle without difficulty large digital<br />
data quantities on low-cost personal computers, and particularly pictures.<br />
So in a lot of urban GIS (geographic information systems) where data<br />
must frequently be updated, it is current practice to work on a basic layer<br />
formed by an orthophotography, that is not costly and therefore can be<br />
often changed. The vector layers are then locally updated, from the<br />
orthophotography, directly by the user himself. This evidently creates an<br />
important market for digital aerial pictures. In addition, since the beginning<br />
of the 1990s many software developments have been made, intended<br />
to do all the photogrammetric restitution directly on PCs, transformed<br />
thus in digital photogrammetric workstations (DPW, described in §2.5) of<br />
lower and lower cost. Users of such DPW are no longer only societies<br />
of photogrammetry, but are more and more technical users coming from<br />
sectors quite external to photogrammetry, who themselves do the restitution<br />
of objects that interest them. The diffusion of the DPW also creates<br />
a demand for digital pictures, demand that is likely to increase for a few<br />
years.<br />
This need for digital pictures is currently satisfied in two ways: by digitization<br />
of traditional silver halide pictures, and by the more and more<br />
frequent use of digital aerial cameras (see §1.5). Pictures obtained by these<br />
cameras have much better performances (radiometric and geometric as<br />
well), but the cameras’ rarity has made it necessary for several years to<br />
use scanners to digitise the new images acquired by traditional aerial<br />
cameras (which are in general very reliable and of quite long life), and<br />
these devices will remain necessary for a long time to allow the use of old<br />
pictures.<br />
We are going to analyse the techniques used by these scanners in order<br />
to understand their shortcomings and their qualities. These materials themselves<br />
are evidently also in regular evolution, but this evolution is notably<br />
slow and the technologies used seem steady enough, probably because<br />
there were no innovations in the domain of sensors applicable to these<br />
devices for some years. We describe here the situation in the year 2000:<br />
there are two classes of scanners, adapted to aerial pictures 24 × 24 cm.<br />
There are those that have specifically been developed for photogrammetry<br />
that are very precise, and thus very costly considering their low diffusion<br />
(Kölbl, 1999), and to the contrary these are scanners of A3 format, less<br />
precise but for a large public, much cheaper, and that can also be used.<br />
The specification of these scanners is as follows: to allow the processing<br />
of aerial pictures 24 × 24 cm, pictures on paper or on film, an analysis<br />
pixel smaller than or equal to 30 m, in colour or in black and white,<br />
with a good geometric precision. The possibility of processing some original<br />
roll film directly is also an important specification for a part of the<br />
market.
60 Michel Kasser<br />
1.8.2 Technology of scanners<br />
1.8.2.1 Sensors<br />
Sensors used by off-the-shelf devices in 2000 are all based on chargecoupled<br />
devices (CCDs, already seen in §1.5), being combined with an ina-plane<br />
analysis of images. Photomultipliers were previously used with an<br />
installation of images on a drum, which was not as easy, but this generation<br />
of equipment is no longer available. These CCD are used in three ways:<br />
in linear CCD, in matrixes, or possibly in TDI linear CCD. Let’s look at<br />
these sensors in detail:<br />
• The linear CCD are used in order to be able to scan at once with a set<br />
of 10,000 or 12,000 detectors, either on only one linear CCD, or while<br />
joining several shorter linear CCD. To get an analysis of colour, one<br />
uses three of these parallel linear CCD juxtaposed, each equipped with<br />
one of the three RGB filters. The group formed by these linear CCD and<br />
the necessary optics is then displaced according to only one axis in relation<br />
to the picture to be digitized, either it moves before the stationary<br />
image, or the inverse. The scan of a whole picture requires several<br />
successive passes, which therefore have to be perfectly connected.<br />
• CCD matrixes used have the order of 2,000 × 2,000 pixels. There are<br />
much larger matrixes, but given their high cost they are not a good choice<br />
as their use would not appreciably accelerate the process. The detector<br />
and its optic move in a sequential way according to two measurements<br />
in front of the image to be processed, in order to cover all the useful surface<br />
(it is necessary according to the size of the chosen digitization pixel<br />
to cover up to 30,000 × 30,000 points of analysis), and there again the<br />
mechanical displacement must permit an excellent connection between<br />
the individual pictures that will make up the whole image.<br />
• The linear CCD functioning according in TDI mode (TDI for ‘time<br />
delay integration’, an inauspicious enough acronym because it explains<br />
nothing). One exploits an oblong matrix (for example 2,048 × 96<br />
pixels), and one uses it like a linear CCD (here of 2,048 pixels). While<br />
making it advance regularly during the digitization as if it were a<br />
normal linear CCD, one transfers charges produced from a line to the<br />
following one (here 96 times), precisely at the same speed as the<br />
displacement of the picture in the plane of the CCD. Thus a whole<br />
set of charges recovered at the exit has been generated by only one<br />
point of the analysed picture. So these charges have been produced by<br />
many successive detectors (here 96), which homogenises the responses<br />
of all detectors to a remarkable degree. It is similar to the way of<br />
working that is described in §1.5 to perform the forward motion<br />
compensation on matrix airborne cameras. One gets therefore a very<br />
regular response from all detectors during the scan.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Scanners for aerial image digitization 61<br />
These devices present different advantages. For example, the time of integration<br />
of a pixel facing the analysed photograph will be much longer for<br />
a matrix than for a simple linear CCD, which will permit one to use a<br />
less powerful lighting to illuminate the film, and will prevent undesirable<br />
thermal effects due to the powerful lamps. The analysis of the colour by<br />
three neighbouring linear CCD equipped with filters is simpler to achieve<br />
than with a matrix, for which it is necessary to commute a set of filters<br />
sequentially before every acquisition of an elementary zone. On the other<br />
hand, if filters are put down directly on linear CCD, which is generally<br />
the case, it is not possible to modify them, and there is no means to regulate<br />
differently the lighting of each channel, which is possible with a matrix<br />
system. Obviously it is necessary that the focusing of the optics be carefully<br />
tuned, and with linear CCD systems it must be the same for the three<br />
linear CCDs, which is not easy considering differences of wavelengths and<br />
therefore of diffraction effects, which are very appreciable for extremely<br />
small pixel sizes.<br />
Let us note that irregularities of detector radiometric response may obviously<br />
be corrected by a preliminary calibration, but that those that subsist<br />
(dust particles, calibration errors) will create periodic and annoying artefacts:<br />
• for linear CCD, some parallel strips will be displayed on the whole<br />
document, among which one will also note those due to the successive<br />
passes of linear CCDs;<br />
• for matrixes it will be due to the repetition of errors according to a<br />
regular paving, and of radiometric discontinuities between successive<br />
positions of the matrix. These shortcomings will be generally far less<br />
spectacular than for linear CCD.<br />
Besides, behind every CCD there are special electronics (voltage-current<br />
amplifier, digital to analogue converter) that must be optimized so that the<br />
reading noise of the CCD is as low as possible, with a digital sampling<br />
ranging from 8 to 12 bits. But this very high quality sampling is not so necessary<br />
considering the noise of the photographic process itself, which does<br />
not justify more than 6 useful bits. One will find normally all classic defects<br />
of the CCD (echoes, blooming . . .) that are described in the corresponding<br />
technical literature (Kodak, Dalsa, Phillips, Toshiba, Thomson . . .).<br />
1.8.2.2 Mechanical conception<br />
To displace an optic as a whole and to reposition it elsewhere within some<br />
microns is not a simple operation. It is also one of the major technological<br />
difficulties of scanners. That one uses linear CCD or matrixes doesn’t<br />
create major differences in this respect. It is therefore mainly problems<br />
of mechanics, that are generally solved using mechanisms having some<br />
extremely reduced play, achieved in materials only giving a weak coefficient
62 Michel Kasser<br />
of differential thermal dilation between the various sub-assemblies: one<br />
does require from this mechanics only that it be extremely reproducible<br />
in its errors. One proceeds to a very tidy calibration of the whole while<br />
observing a regular and known pattern with a very high precision (patterns<br />
on glass plate, known to about 1 m). One then deduces a set of correction<br />
parameters to apply to the different elementary pictures assembly in<br />
order to reach a precision compatible with the specifications (generally,<br />
some m). But it is necessary to be very prudent with the ageing of these<br />
structures, because of unexpected effects of dust on movements, or because<br />
of any play that may develop and compromise the validity of tables of<br />
correction parameters as well.<br />
1.8.3 Size of analysis pixel<br />
The available pixel sizes are very variable, starting at 4 m on the best<br />
resolution devices up to more than 300 m when one achieves a coarse<br />
sampling. Generally these large pixel sizes are reached by regroupings of<br />
data obtained on the smallest possible analysis pixel. This generalized use<br />
of a very small pixel and its regrouping by software is also used to rectify<br />
the geometry of under-pictures between them, which without re-samplings<br />
are still very expensive in time calculation and in memory space.<br />
Recommended sizes of pixel evidently depend on the quality of the image<br />
to be digitized and on the photogrammetric work to be performed. In<br />
§1.9, where Christian Thom studies problems bound to the signal/noise<br />
ratio of photographic emulsion, one will see that the smaller the pixel, the<br />
more the signal/noise ratio deteriorates, and this well beyond the possible<br />
defects of the scanner. A size of 30 m is satisfactory in most of the<br />
studies of aerotriangulation type, DTM, orthophotography, etc. If one uses<br />
smaller sizes (e.g. 15 m), gains in precision are often modest (Baltsavias,<br />
1998) but at a high cost in terms of computer complications.<br />
1.8.4 Radiometric quality, measure of colours<br />
Considering the quite mediocre radiometric quality of photographs (with<br />
a fidelity to the original object that is weak considering the chemical process<br />
used), it is certain that there are not many conceptual difficulties in<br />
achieving an analysis of images that practically does not bring any data<br />
deterioration. In particular the restitution of colours, in such conditions,<br />
looks more to the operator’s artistic sense (to provide a document satisfying<br />
the customer’s eye) than to a rigorous respect of physical laws. The<br />
available devices offer a large palette of tools in order to best balance the<br />
RGB components. Let us note, however, that some mechanical defects can<br />
drive to a small geometric colour component shift, which will be able to<br />
generate some local artefacts especially visible on contrasted objects with<br />
straight edges, linear or of a size close to the pixel size.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
1.8.5 Large A3 scanners<br />
Scanners to the format A3 are available on the market, which are<br />
therefore extensively capable of digitizing a whole aerial photograph,<br />
and intended for the technical public in general. They are incomparably<br />
cheaper than the specialized devices for photogrammetrists (around<br />
50 times less) and it is interesting to note their possibilities of use.<br />
Those available have to the date of writing, sizes of active pixel from<br />
30 to 40 m, using a triple linear CCD with RGB filters, which permits<br />
a digitization in only one pass and therefore a good geometric homogeneity,<br />
some geometric errors ranging to 2 pixels that generally create<br />
an affine distortion. It is to be noted that such distortions are also typical<br />
of films, because of processes that they undergo at the time of the<br />
development and drying, and that these errors are systematically modelled<br />
by one additional unknown in phases of aerotriangulation. It is therefore<br />
possible to use such scanners for certain photogrammetric studies<br />
not requiring the maximal precision, and a fine assessment of their<br />
limits and the temporal stabilities of error models is to be done for each<br />
of them.<br />
References<br />
Radiometric and geometric precision 63<br />
Kölbl O. (1999) Reproduction of colour and of picture sharpness with photogrammetric<br />
scanners, findings of OEEPE tea scanner test, Proceedings of tea OEEPE<br />
Workshop on Automation in <strong>Digital</strong> Photogrammetric Production, Marne-<br />
La-Vallée 22–24 June, pp. 135–150.<br />
Baltsavias E.P. (1998) Photogrammetric film scanners, GIM, vol. 12, no. 7, July,<br />
pp. 55–61.<br />
1.9 RELATIONS BETWEEN RADIOMETRIC AND<br />
GEOMETRIC PRECISION IN DIGITAL IMAGERY<br />
Christian Thom<br />
1.9.1 Introduction<br />
One of the first problems which the digital picture user is confronted with<br />
is the size of the pixel that he is going to use for his scan. Intuitively, he<br />
feels sure that, the smaller the step of the scan, the better will be the precision,<br />
at least relatively, of the result, at the cost unfortunately of a greater<br />
data volume. But generally he does not know that this better resolution<br />
will also be paid for, as we shall see, in terms of radiometric quality. This<br />
loss of radiometric precision also has an indirect but certain consequence<br />
on the geometric precision. One will understand this, according to the<br />
principle that ‘who can do more, can do less’ (one can always gather some
64 Christian Thom<br />
small pixels to build bigger ones . . .). Nevertheless, one does not risk<br />
getting a poorer result, paradoxically with small pixels than with big ones.<br />
But as the size finally chosen is still the result of a compromise between<br />
the geometric precision and the cost/volume of data/time processing,<br />
it is clear that one must take into account the effect of the radiometry on<br />
geometry to find the best compromise.<br />
What is true for digitized pictures is true for digital pictures, that is to<br />
say those provided by digital cameras, and this is because of two points.<br />
First, a better geometric resolution itself has a cost, because it intervenes<br />
directly, and fairly linearly, on the cost of the image acquisition: it is not<br />
a question of a simple adjustment of the scanner, but is rather like the<br />
choice of the scale in the classic case. Second, digital picture radiometric<br />
quality is much better than that of the digitized pictures (under usual conditions<br />
of digitization). One may easily understand that the aforementioned<br />
effect is more crucial in this case.<br />
A poor radiometry has obvious effects on other aspects than the<br />
geometric precision, for example aesthetics and the interpretability of<br />
the picture; they go far beyond the scope of this chapter because they are<br />
too subjective to be modelled in a rigorous manner. Yet, they are not<br />
completely different because the geometric precision that one can reach<br />
in a picture certainly has an influence on its interpretability, and maybe<br />
on its aesthetics.<br />
1.9.2 Radiometric precision<br />
This notion is relatively simple. Radiometry is the measure of the energies<br />
given out by the objects of a scene. Its precision is therefore very definite,<br />
as for any measure. However, in the context of this survey, the radiometry<br />
in itself is still not accessible to us. Indeed, if the digital picture sensors are<br />
almost always radiometric, this is not the case with the digitized film, where<br />
the digital data of the picture are functions of the radiometry, monotonicgrowing<br />
and limited, but not well known, and that, besides, depend on the<br />
conditions of development of the emulsion. Fortunately, it will only have<br />
little impact on our analysis, because we try to value the precision of<br />
localization of visible detail in the images, and this precision will always be<br />
a function of the ratio between the noise and the contrast of the detail. This<br />
ratio may be valued more or less directly in the image.<br />
If the notion is simple, its evaluation is not, especially in the case of<br />
the digitized pictures. If models of noise are well known for the digital<br />
sensors, they are not in the case of digitization, where the noise of<br />
the scanner, noises bound to the emulsion and its process superimpose<br />
themselves. Here, we will use data that we acquired from experiments (1),<br />
and from measures on the digital camera of IGN-F (2). However, it is<br />
necessary to investigate the origin of these noises, because some aspects<br />
have an impact on our topics.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Radiometric and geometric precision 65<br />
1.9.2.1 <strong>Digital</strong> camera<br />
The noise model here is quite simple. The sources of noises are:<br />
1 the reading noise of the sensor, due to the electronics of sampling,<br />
essentially Gaussian;<br />
2 the photon noise, due to the corpuscular nature of light, according to<br />
a law of Poisson;<br />
3 differences of pixel sensitivity (may be corrected by calibration);<br />
4 differences of dark current (may be corrected also by calibration).<br />
In general, it is the photon noise that dominates. It grows according to<br />
the square root of the signal, and its value depends on the total number<br />
of photons that every pixel can contain. For example, in the case of the<br />
KODAK sensor used at IGN-F, this number is 85,000, and the signal/noise<br />
ratio at full capacity is 300.<br />
1.9.2.2 Digitized pictures<br />
Sources of noises are here more numerous and not very well known:<br />
1 noise coming from the sampling unit, which depends on the equipment<br />
used, the size of pixel, the level of grey obtained, etc.;<br />
2 noise coming from the emulsion.<br />
Let us look at this second point in more detail. One knows that the photographic<br />
process is based on detection of light by ions of silver, then to an<br />
amplification of the signal by the developer, that makes every nucleus<br />
created by a photon detected in a grain of silver of a certain size grow<br />
(some microns). The film is then analysed by the scanner that values its<br />
optic density, that is to say the proportion of light that passes between<br />
grains of silver. One immediately sees numerous consequences:<br />
• The response of the film is not linear. Indeed, the photochemical process<br />
of grain production is not linear, because several photons are necessary<br />
so that a grain can be created, producing low sensitivity of the<br />
film to low illuminations. But even though one may suppose this<br />
process as linear, when the density of silver grains grows, they make<br />
‘shades’, when a grain overlaps another. The marginal sensitivity for<br />
a proportion of R grain will thus be 1 R. This gives us a function<br />
according to the illumination i of R(i) 1 exp (ki). The optic<br />
density, which is proportional to the logarithm of 1 R, is therefore<br />
also more or less proportional to the illumination of the emulsion.<br />
• If one considers the middle of the dynamic range of the film (R <br />
1/2), for a size S of the pixel analysis and a characteristic size of grain<br />
of s, if i expresses the number of photons detected, one has:
66 Christian Thom<br />
Surface of the pixel: S 2<br />
Surface of the grain: s 2<br />
For small values of i one notes that k s 2 /S 2<br />
For R 1 ⁄2, i log (1/R)/k log (2) S 2 /s 2<br />
The signal to noise ratio is<br />
r √i √(log 2) * S/s.<br />
(1.33)<br />
Let us take, for example, the following values: S 20 m, s 1 m:<br />
one finds r 16.6. Recall that under similar conditions, a sensor<br />
directly digital may benefit from a signal/noise ratio of about 150.<br />
This gives an indication of the problem . . .<br />
If we consider the values obtained, it seems justifiable to neglect in<br />
what follows the scanner’s own noise in as much as we don’t know<br />
it, and as it depends on the type of device used.<br />
One also notices in this formula that r is proportional to the dimension<br />
of the analysis spot. In general the spot of analysis on a scanner<br />
is bigger than the step of sampling to avoid any aliasing, which slightly<br />
improves the previous values, at the cost of a poorer MTF (modulation<br />
transfer function) of course.<br />
Finally, this formula gives us an idea of the ratio between the<br />
quantum efficiency of the film and of the CCD sensors. Indeed, our<br />
experience shows us that conditions of image acquisition (exposure<br />
time, relative aperture) in the two systems are practically the same.<br />
One sees that on a pixel of 20 m, the film detects 280 photons,<br />
whereas on a pixel of 9 m, the CCD detects about 20,000 of them,<br />
which corresponds to a ratio of 350 in sensitivity . . .<br />
These very simple equations explain why the CCD sensors can be smaller<br />
than the usual focal planes of classical cameras, and that their pixels are<br />
themselves smaller, but that they retain nevertheless an enormous advantage<br />
in terms of radiometric quality. The entire problem is to exploit it to<br />
the fullest, to compensate their relative apparent lack of resolution.<br />
1.9.3 The geometric accuracy<br />
We will concern ourselves here in the geometric accuracy in the picture,<br />
and not in the accuracy of objects in the scene. This means that we will<br />
make abstraction of all problems of systematism bound to the image acquisition,<br />
even if they may be important (e.g. the aerial cameras based on<br />
linear CCD sensors), where the poor knowledge of parameters of external<br />
orientation of the sensor leads for every line of the picture to a poor precision<br />
of pixel localization.<br />
One can distinguish two different ideas in this domain. First of all, the<br />
precision of positioning of a detail in the picture, this detail being able to
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
be punctual, linear, or area, this one going back to the linear case in<br />
general because we position its limits most of the time. In the same way<br />
the punctual is often the intersection of two linear details (corners of<br />
building, for example), although there are also some merely punctual<br />
details. The fact that the detail is distributed in general on several pixels<br />
obviously helps its localization, because one can have several evaluations<br />
of the measured position, and therefore take the mean value.<br />
Then, there is the more delicate problem of the separation of two nearby<br />
details, or resolution, which affects the interpretability of the picture more<br />
than the precision itself, but which we mention here because it is more<br />
demanding in terms of pixel size. Indeed, it is clear that even with pictures<br />
of very good quality it is impossible to separate details less than two pixels<br />
apart, whereas in terms of localization precision, one can without difficulty,<br />
in the same conditions, aim at a few tenths of a pixel. If the size of<br />
the pixel is evidently an element determinant of this precision, it is not<br />
the only one. The resolution of the optics is another. All situations can<br />
present themselves between two extremes:<br />
• fuzzy pictures over-sampled with pixels which are too small;<br />
• high-resolution pictures under-sampled with pixels which are too large.<br />
Both cases can be found for example with a scanner either poorly set up<br />
or not tuned, and the second with digital cameras not having adequate<br />
optics. These situations evidently include the ideal case of pictures sampled<br />
while respecting the limit of Shannon, but it is very rare, because the characteristics<br />
of optics are not constant in their whole field of view, and<br />
therefore a regular sampling does not permit the criteria of Shannon to<br />
be respected everywhere.<br />
One is confronted therefore, in general, often with no ideal and sometimes<br />
with unknown situations. We will examine some simple typical cases,<br />
in general at the extreme ranges of the possible, the intermediate situations<br />
being always more complicated.<br />
1.9.3.1 Precision of localization<br />
Radiometric and geometric precision 67<br />
The case of under-sampled pictures<br />
With regard to the punctual details, the precision is easy to determine:<br />
there is no means to determine the position of the detail inside its pixel<br />
(see Figure 1.9.1). We deduce a RMS precision in x and y of 0.29 pixel,<br />
independent of radiometric precision. In the case of a linear detail, the<br />
precision that one can get depends on the a priori knowledge one has on<br />
the nature of the detail (is it straight?), and of its orientation in relation<br />
to the axes of the picture. The worst case is evidently a detail parallel to<br />
one of the axes of the picture, and having the same precision of 0.29 RMS
68 Christian Thom<br />
pixel. On the other hand, for a differently oriented detail, the distribution<br />
of the energy in the different pixels crossed by the detail permits the precision<br />
to be improved, which depends then on the characteristic dimension<br />
of the detail by which one can consider it as straight. Since one uses the<br />
value of pixels crossed, radiometric precision will have an importance, but<br />
this is difficult to evaluate. With regard to area details, the problem is to<br />
position the edge of the surface. This situation, fortunately more frequent<br />
than the previous (fields, roads, buildings, etc.), is more favourable, because<br />
sub-pixel position of the edge is function of the quantity of energy received<br />
by the pixel containing the edge. If we simplify the problem taking an<br />
edge parallel to the y axis, of c contrast, one understands that its subpixel<br />
position in x can be calculated by: dx dg/c, dg being the difference<br />
of radiometry in relation to the pixel of reference. The precision on x is<br />
therefore directly bound to the radiometric noise. To give an idea, with a<br />
contrast of 1/10 normalized to the value of saturation of the sensor, one<br />
gets in the case of the film digitized at 20 m a noise with a rms of 0.6<br />
pixel, which means that no profit is possible, and in the case of the digital<br />
sensor 1/15 0.07 pixel, which is substantial. Let us push a little farther<br />
the problem of the film. Indeed, why use smaller pixels to the digitization?<br />
Radiometry<br />
In input<br />
Values of<br />
pixels<br />
Figure 1.9.1 Limit of zone, under-sampled situation.<br />
C<br />
dx<br />
x
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Radiometric and geometric precision 69<br />
The noise grows as 1/S, and therefore if one is only interested in a pixel,<br />
the precision will be the same, whatever the value of S. Fortunately, it is<br />
necessary to take into account the fact that for a unit of length of<br />
the detail, one can take the mean value of the position of the detail as<br />
many times that it contains pixels. One will have therefore a precision<br />
proportional to s.(S/L) 1/2 /c, where L is the usable dimension of the<br />
detail. In fact one cannot decrease the size of the pixel indefinitely without<br />
reaching the limit of resolution of the optics, and therefore moving to<br />
the next case.<br />
The case of over-sampled situations<br />
We study here the case where any transition of radiometry is progressive.<br />
All detail is represented therefore on several pixels. There, all the previous<br />
cases are reduced to the case of one edge. Indeed, a linear detail is described<br />
by two successive edges, their characteristic dimension being, however, the<br />
half of one edge (the transition of an edge is in fact the integral of that<br />
of a linear detail, i.e. the point spreading function (PSF), projected on the<br />
axis perpendicular to the detail). We will then look at the localization of<br />
a zone of constant gradient of width D and useful length L. The D size<br />
is bound to the width of the PSF of the optics. One can consider that it<br />
is the half-maximum width of this one in the case of the side of an area<br />
detail, and of one half of this one for a linear detail. Our model here is<br />
very simple, but it will be nevertheless sufficient to analyse the problem<br />
in a qualitative manner (see Figures 1.9.2 and 1.9.3).<br />
Radiometry<br />
L<br />
C<br />
Figure 1.9.2 Limit of zone, over-sampled case.<br />
S<br />
D
70 Christian Thom<br />
Radiometry<br />
L<br />
The precise position (sub-pixel) in x is given by the mean value of the<br />
zone where the gradient is constant, divided by this gradient:<br />
g<br />
Dx <br />
LD<br />
The noise on this value is characterized by:<br />
x g<br />
S<br />
√LD<br />
C<br />
Figure 1.9.3 Linear detail, over-sampled case.<br />
S<br />
D g<br />
S2 <br />
C L S2<br />
C<br />
D<br />
C g S<br />
C D<br />
L .<br />
. (1.34)<br />
(1.35)<br />
Here again, in the case of a digitized image, the size of pixel S also intervenes<br />
into the denominator, and therefore is independent of this one, but<br />
is a function of the grain size.<br />
It is necessary to note that one should not overuse the L parameter.<br />
Indeed, it is rare that one can suppose that a detail is linear on an important<br />
length. Besides, we assume we know its orientation, which is not in<br />
general the case, and introduces a supplementary unknown, and therefore<br />
more imprecision.<br />
Let us recall that in the case of the linear detail, the assessment of the<br />
position can also be done with the other side of the detail, and that D is<br />
one half of the previous case. Thus the benefit is in relation to the edge<br />
length, a factor 2 at the end.<br />
D
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
General case<br />
In fact, in practice one is most often confronted with intermediate situations,<br />
where the PSF of the optics is the order of the pixel, or of a pixel<br />
fraction, and is not known with precision, or varies one point to another,<br />
etc. The methods of precise positioning described may give in this case<br />
mediocre results with residual bias. This is due to the fact that the PSF is<br />
itself poorly sampled. Yet these methods always bring an improvement in<br />
relation to the situation seen in the case of under-sampled pictures whose<br />
results can be considered therefore to contain major errors.<br />
Conclusion<br />
One can extricate several findings from what we have just seen:<br />
• to work with well-sampled pictures is important if one wants to fully<br />
benefit from this kind of technique;<br />
• however, to position the edge of area details, a poor sampling is sufficient;<br />
• the use of this kind of technique on the digitized images is useless,<br />
except for very contrasted details, the noise in this type of pictures<br />
being far too dominant.<br />
1.9.3.2 Spatial resolution<br />
Radiometric and geometric precision 71<br />
Thin detail detection<br />
Contrary to what one commonly believes, it is not impossible to see in a<br />
picture details smaller than the size of the pixel. Just note that it is sufficient<br />
that the contrast of the detail, multiplied by the ratio of the surface<br />
occupied by the detail to that of the pixel, is in correct relation with the<br />
radiometric noise.<br />
An example of this property is given in Figures 1.9.4 and 1.9.5. These<br />
are pictures taken on the same zone, a pond over which passes several<br />
high-tension power lines. The first image is an excerpt of a picture coming<br />
from one digital camera of IGN-F. The second image is a digitized image.<br />
Sizes of ground pixel are similar, around 75 cm. On these two excerpts,<br />
one distinguishes cables of the thickest line (probably doubled), but the<br />
finest line is visible only on the picture from the digital camera. The section<br />
of cables is 3 cm, that is to say of the order of the 1/20th of the pixel<br />
size! One sees that here also, a good signal to noise ratio in pictures allows<br />
one to compensate a poor geometric resolution.
Figure 1.9.4 IGN-F digital camera.<br />
Figure 1.9.5 Digitized image.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Radiometric and geometric precision 73<br />
Resolving power<br />
It seems at first that radiometric quality is not able in any way to compensate<br />
for a poor geometric resolution. Indeed, two details situated in the<br />
same pixel will never be able to be separated. This proposition is obvious,<br />
but it is necessary to state it however.<br />
One can view the problem in the opposite direction, that is to say to<br />
interrogate oneself on the real resolving power of pictures whose radiometric<br />
quality is insufficient. Traditionally one is supposed to discern<br />
two details (choose for example bright ones) if they are separated in the<br />
picture by dark pixels. It is clear that in the case of details poorly<br />
contrasted, and in pictures of poor quality, the bright pixels because of<br />
the presence of details will have a certain chance of being darkened<br />
by the noise, and the dark pixel that separates them will have a possibility,<br />
to the contrary, to be brighter. The inverse situation may happen,<br />
where an object will be split falsely. One sees that one rejoins here the<br />
problem of segmentation, since one poses the question: ‘Does this pixel<br />
belong to this object?’<br />
Let’s study a concrete case, for example: the segmentation of a supposed<br />
homogeneous zone. We deliberately choose a simple algorithm, where a<br />
step on radiometry determines the assignment to one zone or to another.<br />
If the contrast between the two zones is c, the step will be placed to c/2<br />
(one supposes the noise to be independent of radiometry). What is, in these<br />
conditions, the probability that a pixel is poorly classified? It is a simple<br />
problem of statistics, bound to the interval of confidence. The success of<br />
the segmentation process and its precision will obviously depend on this<br />
parameter, as a non-trivial function of the algorithm used, and a priori<br />
knowledge that one has from structures in the picture (straight edges for<br />
example).<br />
One can see in Figures 1.9.6 and 1.9.7 the same zone in two simultaneous<br />
image acquisitions, one with a digital camera of IGN-F, the other<br />
with a traditional camera. If one pays attention to the ridges of roofs, one<br />
notices that they are clearly delimited in the digital picture, whereas some<br />
edges are indiscernible on the digitized picture. This loss of visibility will<br />
evidently have consequences on the precision of its restitution, notably<br />
with automatic means. It is unfortunately impossible to quantify this here,<br />
because it evidently depends on the algorithms used.<br />
In Figures 1.9.8 and 1.9.9 one can see the result of the filter of extraction<br />
of contour from Adobe Photoshop for example. It clearly appears<br />
that the extraction of roof edges can be performed more easily and<br />
more precisely on the digital picture. Note the side of roads in the shady<br />
regions.<br />
To illustrate even better the influence of radiometric quality, and<br />
its interactions with geometry, see Figures 1.9.10 and 1.9.11: the same<br />
picture is presented here, but after it has been under-sampled by a
Figure 1.9.6 IGN-F digital camera.<br />
Figure 1.9.7 Scanned image.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 1.9.8 IGN-F digital camera, contours.<br />
Figure 1.9.9 Scanned image, contours.
76 Christian Thom<br />
Figure 1.9.10<br />
Under-sampled scanned picture.<br />
factor 2. This operation, as it averages pixels, decreases the noise, and<br />
one sees that the process of contour extraction now works better, but its<br />
result will be evidently less precise, being obtained with a pixel twice as<br />
large.<br />
We didn’t mention here the question of the sampling technology, but it<br />
has an influence of course. When the PSF is larger than the pixel, the<br />
contrast between two nearby details tends to soften, and therefore the<br />
restitution of the two details will be sensitive to the noise. When it is<br />
smaller, one doesn’t note an appreciable effect.<br />
Conclusion<br />
It is clear that for what concerns the geometric resolution, radiometric<br />
quality brings much with regard to the detection of details. It has an influence<br />
on the separation in the case of weakly contrasted details, that is to<br />
say whose contrast is the same level as the noise present in the picture. It<br />
is necessary to mention that these details are frequent in urban zones,<br />
materials of construction often being of similar albedo (different types of<br />
coating, for example), and the frequent shade zones.<br />
1.9.4 General conclusion<br />
Figure 1.9.11<br />
Contours from Figure 1.9.10.<br />
The impact of radiometric quality on the geometric quality is still not easy<br />
to evaluate. The problem is complicated by the fact that it depends on<br />
algorithms used for the restitution, and on the adequacy of the picture<br />
sampling. Yet, it is real in most cases. The comparative studies conducted<br />
on the digital camera of the IGN-F and on digitized pictures, to scales<br />
ranging from 1/15,000th to 1/30,000th, clearly show that the very large<br />
available dynamics of the CCD camera allows one to reach the same uses<br />
as the digitized silver halide picture, this in spite of a pixel of a size about
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Radiometric and geometric precision 77<br />
twice as large. But this coefficient of 2 is only valid in this range of scales,<br />
and other studies should specify its value for smaller pixels: this coefficient<br />
doesn’t depend indeed only on radiometric quality, but also on the<br />
type of object being surveyed.
2 Techniques for plotting<br />
digital images<br />
INTRODUCTION<br />
This section deals with the ways of extracting the 3D data from the images.<br />
Some considerations, here also, are not really specific to the use of digital<br />
images, but they must nevertheless be examined within the scope of digital<br />
photogrammetry, as in the case of aerotriangulation, which now offers<br />
some new aspects with digital images (such as the automatic extraction of<br />
link points, for example). We will start with a summary of the image<br />
processing techniques seen from the point of view of the photogrammetrists,<br />
i.e. when the geometry is critical (§2.1). In the same way, the techniques<br />
of data compression are analysed for the specific case of digital images<br />
whose geometry, here too, is critical (§2.2). A summary about the use of<br />
GPS in airplanes (§2.3.1) is then presented: this is not specific to digital<br />
photogrammetry, but it deals with many practical problems of today’s<br />
photogrammetry. Then a short presentation about the aerotriangulation<br />
process is proposed (§2.3.2). Then, under the title ‘automatization of the<br />
aerotriangulation’ is presented the automatic extraction of link points, and<br />
their use to compute the index maps (§2.4), which is the first step when<br />
processing a set of images. Then we close this chapter with a survey of<br />
the techniques used for digital photogrammetric workstations (§2.5). The<br />
automatic cue extraction methods, available at the time of completing this<br />
book, for planimetric items (the vegetation, the buildings, the roads, . . .)<br />
are still very uncertain and cannot be yet considered as operational, thus<br />
we have not presented them, as they are still in the research laboratories.<br />
Of course in Chapter 3, the automatic extraction of the altitude will be<br />
explained, with all the induced problems of DTM, DEM, DSM: this is the<br />
only operational process available that uses all the possibilities offered by<br />
digital images (and it is the simplest one, any other cue extraction requires<br />
much more human interpretation from the operator, and thus is also much<br />
more difficult to implement on a computer).
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
2.1 IMAGE IMPROVEMENTS<br />
Alain Dupéret<br />
2.1.1 Manipulation of histogram<br />
The improvement of the image contrast is a process that tends to make<br />
the images more beautiful to use, on sometimes subjective criteria. Even<br />
if we cannot necessarily say how a good image should look, it is often<br />
possible to say if it is optimized, sufficiently, for example, to see some<br />
details. There are therefore several techniques of image improvement, but<br />
manipulations of a histogram probably represents one of the most used<br />
methods.<br />
2.1.1.1 Definition of the histogram of a image<br />
In statistics, the histogram is a bidimensional representation of the h function,<br />
called density of frequency, defined by:<br />
h(x) h i<br />
if x ∈[a i1, a i]<br />
Image improvements 79<br />
∞<br />
∞<br />
h(x) 0 if x < a0 or if x > ak verifying h(x) dx 1.<br />
(2.1)<br />
In image processing, the histogram is therefore the diagram representing<br />
the statistical distribution (h i, n i ) where h i represents the ith level of possible<br />
colour level among the N possible values and n i the number of pixels<br />
presenting value h i in the digital image. By misuse of language, the h function<br />
is often called a histogram. This function is always normalized in<br />
order to be used like a law expressing the probability that a colour level<br />
n i has to be present in the whole image. The grey level is considered therefore<br />
like a random variable that takes its values in a set of size N, being<br />
often 256 by reason of the usual coding of the elementary information of<br />
every channel on a byte. The new digital sensors, whose intrinsic dynamics<br />
ranges over several thousand levels require more space for storage.<br />
A priori, no hypothesis can be proposed on the shape of a histogram<br />
and it is quite normal to have a multimodal histogram, presenting several<br />
attempts. A corresponding histogram to a Gaussian distribution is possible<br />
but more current in the case of satellite images. The only certainty concerns<br />
the extent, of which it is recommended that it is as large as possible, up<br />
to possibly using the 256 available levels. (See Figure 2.1.1.)<br />
2.1.1.2 Visual representation of data<br />
In practice, when data present a weak dynamic at the level of making their<br />
exploitation difficult, transformations intended to improve the contrast are
80 Alain Dupéret<br />
Figure 2.1.1 Excerpt of an aerial image and the histogram associated with this<br />
image.<br />
still possible and will be presented. These will act by modifying coded<br />
values of the image, either while adapting the transcodage table (Look-<br />
Up-Table) used by the screen for the colour display. The user should keep<br />
in the mind also, that when an image is displayed on a computer screen,<br />
its contrast also depends on the adjustment of the monitor but also and<br />
especially on the spectral sensitivity of the eye, that presents a maximum<br />
for green-yellow colour levels; for a given wavelength variation of d, it<br />
implies a weaker colour level perception for the red or the blue-green, situated<br />
on the edges of the curve of spectral sensitivity of the eye, which is<br />
variable from one observer to another. Other factors are to be taken into<br />
account also, such as the prolonged observation of an intense colour<br />
provoking a transient colour aberration due to the retinal pigments, the<br />
global environment of the point observed that can make appear a more<br />
or less clear colour, as well as side effects.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Image improvements 81<br />
2.1.1.3 Generalities on the modification of histogram<br />
Manipulations of histograms are punctual operations that transform an<br />
input image I(x, y) into an output image I′(x, y) according to a law I′(x,<br />
y) f [I(x, y)]. The f function is presented as a graph for which abscissas<br />
are colour levels of I, and the ordinates colour levels of I′. They are often<br />
destined to fit the dynamics of input images that are initially not contrasted<br />
enough, too clear or too dark. This redistribution of colour levels of the<br />
image can lead to some interpretation errors when the f function is nonmonotonous<br />
or discontinuous:<br />
• the horizontal sections in the graph of f lead to information losses;<br />
• some vertical jumps provoke discontinuities resulting in false contours<br />
or aberrant isolated points;<br />
• some negative slopes lead to local inversions of contrast.<br />
The histogram of the image I′ is centered in ac b and its width multiplied<br />
by a. In image processing software, the parameter ‘brightness’ can<br />
be associated to the b value and the parameter ‘contrast’ to the value a,<br />
slope of the curve transforming Di into Di′. (See Figure 2.1.2.)<br />
2.1.1.4 Dispersion of the dynamics<br />
As the raw image, considered here with pixels coded on one byte, never<br />
takes all possible values between 0 and 255, it results in a contrast default<br />
even though one knows that the eye can only discern a few tens of grey<br />
levels on a screen. Thus it is convenient to spread out the initial dynamics<br />
[m, M] in the largest possible interval [0, 255] with the help of the operation<br />
DI ′ (DI m) × [255/(M m)]. This operation does not provide<br />
n I<br />
b<br />
D I′<br />
c<br />
D I′ = aD I + b<br />
D I<br />
D I<br />
n I′<br />
Figure 2.1.2 Manipulations of histogram.<br />
D I′<br />
Histogram of the output image I′
82 Alain Dupéret<br />
any new information and is more a strategy of display on the screen than<br />
an interpolation technique. It is often proposed at a more elaborate level,<br />
giving better results, by calculating two levels (min and max), as<br />
min<br />
0<br />
h(x)dx h(x)dx , 0.01 or 0.00,<br />
(h represents the function of distribution of colour level levels in the image).<br />
Then the dispersed display ranges from [min, max] to [0, 255].<br />
An interesting variant exists when the input histogram is unimodal and<br />
relatively symmetrical and determines min and max according to the statistical<br />
properties of the colour-level distribution, as illustrated in Figure 2.1.3.<br />
For a multispectral image, the dispersion is achieved independently on<br />
the different channels, at the risk of provoking a modification of the colour<br />
balance, each channel having initially a dynamics that is its own for a<br />
given raw image.<br />
It is quite possible to proceed to a non-linear redispersion of the dynamics<br />
whose advantage is to reinforce contrasts for a part of colour levels of a<br />
histogram: logarithmic shape to reinforce the dark zones, exponential one<br />
for the clear zones. Often used, this strategy has the drawback of decreasing<br />
contrasts in the zones of the histogram not affected by the strong curvature<br />
of the transfer curve but it does allow a more faithful replication of<br />
the way the eye responds to different levels of brightness. (See Figure 2.1.4.)<br />
A strong slope on the central curve means an accentuation of contrasts in<br />
the corresponding zones, which are here the numerous dark parts of the<br />
right image. The thinning of contrasts in the clear parts is not here visually<br />
bothersome.<br />
2.1.1.5 Equalization of a histogram<br />
The equalization of a histogram is a method that looks for an equiprobable<br />
statistical distribution of levels hi , which means to put ps (s) c where<br />
c is a constant.<br />
The relation pr (r) dr ps (s)ds being valid for all values, the search for<br />
T in R T(S) leads therefore to<br />
s<br />
0<br />
r<br />
c ds pr (r) dr ⇒ c 1<br />
0<br />
255<br />
max<br />
s<br />
s 0<br />
p r (x) dx ,<br />
which provides the sought-after relation. (See Figure 2.1.5.)<br />
The previous formulation presumes that r and s can take continuous<br />
values. With the digital images, it may be only a finite number of values in<br />
the interval [r, rdr]. When redistributing this interval on a larger domain,<br />
the principle of conservation of the number of pixels means therefore that
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 2.1.3 For a given raw image on the left, the visual results are given for an interval [min, max] being worth<br />
successively [moy, moy], [moy2, moy2], [moy3, moy3] toward the right.
84 Alain Dupéret<br />
Figure 2.1.4 Non-linear redispersion of the dynamics: on the left, the original<br />
image, on the right, the modified one through the curve in the<br />
centre.<br />
Figure 2.1.5 Application of the method of equalization of the histogram.<br />
only the concerned pixels are redistributed and that the histogram cannot<br />
be virtually flat.<br />
For an image I constituted of n colour level levels quantified on N levels<br />
D I(k), these last are transformed by the law<br />
D I′(k) (N 1) p I D I(k) where p I n I<br />
n ,<br />
n I representing the number of pixels of values D I(k). As the results D I ′(k)<br />
may be decimal, the method implemented in the software rounds off then<br />
to the nearest integer value, which leads to having several D I(k) levels<br />
associated to D I′(k), to which are affected the sum of probabilities of the
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
D I(k) from which it originates, thus justifying the non-uniformity of the<br />
histogram. To palliate this effect, some methods redistribute from pixels<br />
the more represented in colour levels toward the insufficiently represented<br />
neighbouring values, until reaching a balanced population for every quantification<br />
level.<br />
If the number of colour levels in I′ is therefore lower than in I, all<br />
possible levels not being necessarily represented, this technique allows the<br />
best possible dynamics and gives strong contrasts. Nevertheless, it may not<br />
give visually good results, colour levels of objects photographed not necessarily<br />
being suitable for such a hypothesis.<br />
2.1.1.6 Obtaining a histogram of any shape<br />
The operator can very well choose to have the function ps (s) as the final<br />
histogram, which means then<br />
s<br />
0<br />
r<br />
ps(y) dy pr(x) dx ,<br />
0<br />
which permits one to define s according to r.<br />
Generally speaking, the following steps are achieved:<br />
Image improvements 85<br />
• equalization of the original image y 1 g(r) p r(u)du;<br />
• equalization of the desired histogram y 2 f(s) p s(u) dv;<br />
• application of f 1 (function that is calculated easily by tabulation) to<br />
the histogram g(r), or f (s) g(r) ⇒ s f 1 g(r) (the two histograms<br />
being identical).<br />
2.1.1.7 Improvement of local contrast<br />
The previous methods are global manipulations of the image and do not<br />
take into account the possible disparities of contrast or quality that may<br />
exist within any given image. To avoid this inconvenience, techniques of<br />
local improvement of contrast have been developed. In the neighbourhood<br />
of each pixel of size n × m that is going to be displaced progressively on<br />
all pixels of the image, the local histogram is calculated in order to be<br />
able to apply methods previously mentioned, such as the equalization or<br />
the assignment of a shape of a given histogram (e.g. Gaussian shape). As,<br />
each time, only a new column or a new line appears with the transfer of<br />
a pixel in the image, it is possible not to repeat the complete calculation<br />
of the histogram while using calculations of the previous iteration.<br />
r<br />
<br />
0<br />
s<br />
<br />
0
86 Alain Dupéret<br />
Thus, for a given image, the definitive assignment of a colour level for<br />
a pixel requires each time a process of manipulation of the present value<br />
histogram in a neighbourhood of given size.<br />
The statistical process operates on sub-images of variable size and is<br />
going to use the mean intensity and the variance that are two applicable<br />
properties to model the image appearance. For a mean radiometric value<br />
moy(i, j) inside a mean window centred on the pixel of coordinates (i, j),<br />
and a radiometric variance of (i, j) inside the same neighbourhood, the<br />
value of the radiometry in I may be modified in I′<br />
I′(i, j) A × [I(i, j) moy(i, j)] moy(i, j) with<br />
A kM<br />
(i, j)<br />
, scalar value called gain factor (2.2)<br />
M: mean value of the radiometries in the whole image k ∈ ]0, 1[<br />
Figure 2.1.6 Improvement of the local contrast.<br />
(a) In terrestrial photogrammetry, conditions of lighting are sometimes far from optimum.<br />
Effects of over- or under-exposure are decreased by a local improvement of contrast, but<br />
beyond a given threshold, the noise of image becomes unacceptable.<br />
(b) The image on the left presents clear zones in the top-left corner (hot spot) that can be<br />
suppressed while imposing the image in exit on the right to present a Gaussian histogram<br />
inside a window whose size is chosen by an operator.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Image improvements 87<br />
so as to heighten the contrast locally for the weakly contrasted zones, A<br />
being inversely proportional to (i, j).<br />
The size of the processing window is a predominant criterion:<br />
• if small (3 × 3 to 9 × 9), it permits the strongest possible improvement<br />
of the local contrasts, the side effect being to increase the image noise<br />
(see the pair of images on the Egyptian temple of Dendera shown in<br />
Figure 2.1.6);<br />
• if large (50 × 50, 100 × 100 . . .), it allows the averaging of the values<br />
of colour levels on large zones, thus decreasing the disparities of brightness<br />
in the different places of the image, but its local effects are then<br />
less visible (see the pair of images on the zone of Agen).<br />
These methods are often used to improve the similarity of two images<br />
acquired at different dates; in aerial imagery, it allows the reduction of<br />
Figure 2.1.7 Improvement of contrast.<br />
Stereogram of the raw images in the upper part, and of images with local improvement of<br />
contrast on small size windows; even though the increased noise appears unacceptable on<br />
an image, the stereoscopic examination in zones of shade is nevertheless possible and limits<br />
the need to order a new aerial acquisition or the exploitation of other images.
88 Alain Dupéret<br />
disparities of radiometry between two distinct images when they must be<br />
merged into a mosaic, or even an ortho-image.<br />
Beyond the simple process of a unique image, the use of a couple of<br />
stereoscopic images can be very helpful at medium or large scales, especially<br />
when zones of shades embarrass the perception of objects (see Figure<br />
2.1.7).<br />
The application to multi-spectral images (several channels) is more delicate,<br />
as previously seen, processes having to be applied independently to<br />
each channel to provide a coherent result. In the case of colour images,<br />
the representation can be performed in a colorimetric reference with red,<br />
green and blue main components; every coordinate of a pixel along one<br />
of these axes is the colour level it has in the corresponding spectral band.<br />
The set of points corresponding to pixels in this reference forms a cloud<br />
in which one searches the main axes along which data are best distributed.<br />
This is performed using the method of classification in main<br />
components often used in remote sensing.<br />
The operation of local improvement of contrast can be achieved while<br />
passing in the frequency domain and using techniques of homomorphic<br />
filtering.<br />
2.1.2 Filtering – suppression of the noise<br />
2.1.2.1 The noise in the images<br />
The signal in the image is composed of information, superimposed with<br />
noise; in the case of an image, it may be the noise inherent to the digital<br />
sensor used at the time of the acquisition, of the nature of the photographic<br />
document if it is a document that was digitized. Methods already<br />
presented processed every pixel independently of the neighbouring pixel<br />
value, and also have the effect of increasing without distinction noise and<br />
information in the images, which justifies the processes devoted to<br />
decreasing the noise. This noise is formed of pixels or small isolated pixel<br />
groups presenting strong variations with their environment.<br />
In practice, images have a certain spatial redundancy, the neighbouring<br />
pixels often having some colour levels of very close values. If some general<br />
hypotheses on the noise are made (additive, independence of the signal<br />
and greatly uncorrelated), it is possible to achieve operations of filtering<br />
intended to limit the presence of the noise in images while achieving lowpass<br />
filtering, linear or not, in the spatial or frequency domains. The<br />
justification of the low-pass character of the filter is as follows.<br />
Usually, the noise in images is not correlated with the signal level, and<br />
presents therefore a fairly uniform spectral distribution. As for images the<br />
part of low frequencies is more important than that of high frequencies,<br />
a moment happens where the signal is dominated by the noise, from where<br />
comes the idea of threshold beyond a limiting frequency. The elimination
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Spectral density<br />
Image spectrum<br />
F c<br />
Noise spectrum<br />
Frequency<br />
of this part of the spectra of course destroys a certain part of the signal<br />
which results in a image with little texture and fuzzier contours.<br />
This process can by calculation of Fourier transform of the image on<br />
which is applied a process entrenching all the frequencies beyond the<br />
threshold value, then by an inverse Fourier transform. (See Figure 2.1.8.)<br />
There are two general types of noise (see Figure 2.1.9):<br />
• additive noise:<br />
– of pulse type, whose detail is to tamper randomly with some pixel<br />
colour levels to produce a ‘pepper and salt’ effect in images;<br />
– of Gaussian type, if a variation of colour level following a centred<br />
Gaussian probability density is added on colour levels of the image<br />
– this is the most current type;<br />
• multiplicative noise, as for example that resulting from a variation of<br />
illumination in the images.<br />
2.1.2.2 Use of filtering by convolution<br />
Methods of filtering are numerous and are often classified by type: linear<br />
filtering, non-linear, morphological (opening, closing), by equation of diffusion<br />
(isotropic or not), adaptive by coefficients, or adaptive windows.<br />
The use of convolution operators in the spatial domain (the image) transforms<br />
a I(x, y) image into I′(x, y) using a square window of odd dimension<br />
2n1, moved on the image, by a matrix A with coefficients ai, j so that<br />
I′(x, y) ai, j I(xi, yj) where n i, j n .<br />
i, j<br />
Image improvements 89<br />
Figure 2.1.8 Beyond F c , the spectrum of the image is dominated by the noise.<br />
(2.3)<br />
2.1.2.3 Linear filtering in the spatial domain<br />
The simplest filtering consists in finding the average of pixels situated in<br />
a window of given size and to apply it to the central pixel, and to repeat<br />
this operation in every point of the image.
Figure 2.1.9 Types of noise in images.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
The application of a filtering is a question of convolution of an image<br />
by a kernel, assimilated to the mask of the transfer function of the filter,<br />
whose coefficients can be adapted by the user according to the objective.<br />
Examples of kernels for low-pass filters:<br />
1<br />
9 1<br />
1<br />
1<br />
1<br />
16 1<br />
2<br />
1<br />
1<br />
1<br />
1<br />
2<br />
4<br />
2<br />
1<br />
1<br />
1<br />
1<br />
2<br />
1<br />
or filter by the average,<br />
Image improvements 91<br />
or Gaussian filter.<br />
(2.4)<br />
This type of filtering is adapted to the elimination of Gaussian noise more<br />
than that of pulse noise. The way to use the filter by the average can be<br />
optimized while keeping in mind the processes applied to points of the<br />
neighbourhood; for Gaussian filter, it may be decomposed into two unidimensional<br />
filters. These artifices permit the number of operations performed<br />
in order to achieve filtering to be reduced. (See Figure 2.1.10.)<br />
The use of such a filter decreases high spatial frequencies, as strong as<br />
the size of the window is large. One thus should be careful not to tamper<br />
with objects that one wishes to preserve in the image, because their contour<br />
becomes less and less meaningful with the progression of this filtering.<br />
Moreover, these processes generate side effects the larger the size of the<br />
window or the number of applied iterations is more important. This filtering<br />
therefore considerably degrades the contours and makes the image fuzzy,<br />
with some exceptions, as for example when the signal is stationary and<br />
the noise Gaussian.<br />
2.1.2.4 Non-linear filtering in the spatial domain<br />
Linear filtering has the main drawback of processing in the same way the<br />
parts of signal carrying information and noise, which sometimes justifies<br />
the use of non-linear processes, more able to attenuate the aberrant values<br />
of pixels whose colour level is too distant from the neighbouring ones.<br />
Also called filtering of rank, these methods replace the central or current<br />
pixel by one of the values selected from pixels of the Vxy neighbourhood<br />
of it (x, y): pixel of minimal, maximal or median value . . .<br />
This type of method is more robust with regard to the noise of the<br />
image and has the advantage of assigning an existing value without calculating<br />
a new one.
Figure 2.1.10 Gaussian filtering.<br />
On the left, an image digitized at 21 m on an aerial photograph. The homogeneous zones, including noise, contain a random texture<br />
that does not constitute a useful signal. In the centre, the original image after Gaussian filter. If the noise was attenuated, contours also<br />
became fuzzy. On the right, an image of a digital camera from IGN-F: zones are naturally weakly noisy because of the nature of the<br />
sensor and its very good dynamics.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
2.1.2.5 Median filter<br />
In this process, the value assigned to the pixel is the statistical median<br />
value in the filtering window, whose number of elements must therefore<br />
be odd. The median is indeed the best evaluation of the mean value of a<br />
whole sample, less sensitive to aberrant values, and it is supposed here<br />
that they correspond to the extreme values of the histogram inside the<br />
mask of the filter. (See Figure 2.1.11.)<br />
This process allows one to eliminate the isolated values that correspond<br />
to tips of pulse noise and preserve contours of objects more efficiently than<br />
the linear filters. Nevertheless, distortions misled by this filter on objects<br />
increase with the size of the window used. This filtering is therefore especially<br />
advisable when images present a noise of pulse type, and therefore<br />
important variations of weak extent. Some implementations use a mask<br />
in the shape of a cross centred on the current pixel, a shape which is<br />
preferred to the square shape to limit distortions in the image.<br />
2.1.2.6 Adaptive filtering<br />
Filters are said to be adaptive in the following cases:<br />
Image improvements 93<br />
Figure 2.1.11 Median filter.<br />
Use of the median filter on the previously used raw images. It presents as a main defect to<br />
round corners and to smooth lines whose width is less than a half of that of the filter.<br />
• coefficients consider every term of the filter by the average, by a term<br />
that decreases with the similarity between the considered pixel and the<br />
central pixel of the window;<br />
• the window is chosen either in shape or in size.
94 Alain Dupéret<br />
The filter of Nagao, relatively unknown, provides nevertheless very good<br />
results in the case of aerial images, in particular in urban sites. It is structured<br />
in four steps:<br />
1 a window 5 × 5 around the central pixel is divided into nine windows<br />
of nine pixels each;<br />
2 the first two models are each declined in four diagrams deduced from<br />
the one presented by successive rotations of 90°, 180° and 270°;<br />
3 the homogeneity of every window is measured, with the help of an<br />
indicator of the radiometric variance type;<br />
4 the central pixel is replaced by the most homogeneous average within<br />
the nine zones.<br />
The filter obtained strongly smoothes textures without tampering too much<br />
with the contours of objects in the images. (See Figures 2.1.12 and 2.1.13.)<br />
The filter of Nagao has the defect of smoothing the thinnest features of<br />
the image. Some algorithmic variants therefore exist to find the most applicable<br />
indicators of the homogeneity measure. The size of windows used,<br />
the weighting type used and the shape of masks of application can be<br />
adapted empirically in such a way as to get the best filter for the preservation<br />
of the lines, the corners or the suppression of the noise.<br />
Figure 2.1.12 Filter of Nagao, first step.<br />
Figure 2.1.13 Filter of Nagao applied (right) to the left image.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
2.1.2.7 Filtering in the frequency domain<br />
The filter in the frequency domain takes place in three steps:<br />
• calculation of F(u, v) TF(I(x, y)); in general, it is fast Fourier transform<br />
that is used on the base image;<br />
• multiplication by H(u, v), transfer function of a ad-hoc filter: G(u, v)<br />
H(u, v)·F(u, v);<br />
• obtaining the image filtered by inverse Fourier transform of G(u, v).<br />
The time of computation becomes a major parameter that must be considered,<br />
before any use in an operational context. Practically sometimes one<br />
uses, instead of direct and inverse Fourier transform, a square mask whose<br />
coefficients are the discrete elements of the transfer function of the frequency<br />
domain filter. Butterworth filters, low-pass exponential and Gaussian are<br />
the more frequently evoked, even though, for the meantime, only few<br />
applications in digital photogrammetry use them.<br />
Homomorphic filtering is also usable to correct irregularities of lighting<br />
of an object. Note that a I(x, y) image can be written as the product of<br />
e(x, y)·R(x, y), where e is the function of lighting that rather generates<br />
some of the low frequencies at the time of the calculation of Fourier transform,<br />
and R is the reflectance that rather generates some high frequencies<br />
by the nature of objects that composes it.<br />
The two effects are made additives by creating a logarithmic image so<br />
that:<br />
ln (I(x, y)) ln (e(x, y)) ln (r(x, y)). (2.5)<br />
Homomorphic filtering will increase the high frequencies and reduce the<br />
low frequencies to the point of reducing variations of the lighting while<br />
details will be reinforced, thus permitting a better observation in the dark<br />
zones of the image (see Figure 2.1.14). The function of transfer is the<br />
shape<br />
H( x , y ) <br />
with values as s 1, 0 128 and A 10, these parameters being joined<br />
as follows to the H and L parameters and by<br />
L <br />
1<br />
1 exp (s0) A, H 1 A .<br />
1<br />
1 exp (s 2 2<br />
√x y 0)<br />
Image improvements 95<br />
A ,
96 Alain Dupéret<br />
Γ H<br />
Γ L<br />
H(r)<br />
2.1.3 Improvement of contours<br />
Points of contours form the majority basis of the information included in<br />
the images, probably more than textures, which play nevertheless an essential<br />
role, in particular for the perception of the relief. The recognition of<br />
an object is performed most often only from its contours. The difficulty<br />
resides in the subjective notion that the user presumes to define the utility<br />
and the importance of a contour. The multiplicity of contour detectors<br />
results from the variety of the studied images and from the type of application.<br />
2.1.3.1 Methods of the gradient<br />
This first category is based on the use of the first derivative of the luminous<br />
intensity. The most often met operators are constructed below from<br />
models:<br />
1<br />
0 1<br />
0<br />
1<br />
0 0<br />
1<br />
1<br />
1<br />
1<br />
r = (ω x 2 + ωy 2 ) 1/2<br />
Figure 2.1.14 Homomorphic filtering.<br />
0<br />
0<br />
0<br />
1<br />
1<br />
1<br />
5<br />
3<br />
3<br />
5<br />
0<br />
3<br />
5<br />
3<br />
3<br />
Gradient Roberts Prewitt Sobel Kirsh (2.6)<br />
The three last operators have the advantage of being less sensitive to the<br />
noise than the traditional derivative, but unfortunately also create more<br />
drifts. As each are directional, it is usual to construct similar filters to<br />
detect the contours in the eight possible directions.<br />
1<br />
2<br />
1<br />
0<br />
0<br />
0<br />
1<br />
2<br />
1<br />
.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
The convolution by Sobel operators<br />
1<br />
2<br />
1<br />
0<br />
0<br />
0<br />
1<br />
2<br />
1 ,<br />
2<br />
1<br />
0<br />
1<br />
0<br />
1<br />
(that would be to the number of eight by displacing the whole coefficients<br />
by a quarter of a turn for every new filter) allows one to detect zones of<br />
contours in a given direction.<br />
The superposition of the four images obtained with the above-proposed<br />
filters achieves an image of the amplitude of the gradient; for every<br />
pixel I (x, y) , the maximum found in the intermediate images of coordinates<br />
(x, y) is selected. The presence of a contour can be decided in<br />
relation to a threshold on this value. (See Figure 2.1.15.)<br />
It will often be interesting to reduce the noise of an image with methods<br />
like the median filter or the filter of Nagao, which will not prevent the<br />
contours from suffering from an approximate quality if they are too noisy,<br />
thick and non-continuous.<br />
Figure 2.1.15 Sobel filter.<br />
1<br />
0<br />
1<br />
0<br />
1<br />
2 ,<br />
2<br />
1<br />
2<br />
1<br />
0<br />
1 ,<br />
Image improvements 97<br />
0<br />
1<br />
2<br />
1<br />
0<br />
1<br />
2<br />
1 . . .<br />
0
98 Alain Dupéret<br />
2.1.3.2 Methods of passage by zero of the second derivative<br />
This second category is based on the survey of the passage by zero of<br />
second derivatives of the luminous intensity in order to detect contours.<br />
If zeroes of the second derivative constitute a closed network of lines, the<br />
second derivatives are generally very noisy. To palliate this inconvenience,<br />
it is possible to widen the support of the mask used, to select only passages<br />
by zero accompanied by a strong slope of the signal, or again to filter<br />
contours obtained after the detection.<br />
There are several types of Laplacian:<br />
0<br />
1<br />
0<br />
1<br />
4<br />
1<br />
0<br />
1<br />
0<br />
.<br />
(2.7)<br />
After the application of these masks to the image, one detects contour<br />
points by the detection of passages by zero of the image obtained. The<br />
singleness of the passage by zero provides contours whose thickness is<br />
directly one pixel. (See Figure 2.1.16.)<br />
One shows that, if the noise is Gaussian, the most suitable operator is<br />
constituted by the Laplacian of a Gaussian defined by:<br />
G(x, y) 1<br />
1<br />
1<br />
1<br />
1<br />
8<br />
1<br />
2 4 2 x2 y 2<br />
1<br />
1<br />
1<br />
1<br />
2<br />
1<br />
2 exp x2 y 2<br />
. (2.8)<br />
Another method calculates passages by zero of the directional second<br />
derivative in approximating the intensity in a window by a polynomial<br />
whose coefficients are calculated by convoluting images with masks. One<br />
can calculate second derivatives in the direction of the gradient and find<br />
2<br />
2<br />
4<br />
2<br />
2 <br />
1<br />
2<br />
1<br />
Figure 2.1.16 Use of a Laplacian filter (right) on the left image.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
points of contour as passages by zero of these directional second derivatives.<br />
For a point of coordinates (i, j), the coefficients approximating the<br />
intensity in the basis<br />
1, i, j, (i 2 2<br />
3), ij, (j 2 2<br />
3), (i 2 2<br />
3)j, i(j 2 2<br />
3), (i 2 2<br />
3), (j 2 2<br />
3)<br />
are the following:<br />
1<br />
9 1<br />
1<br />
1<br />
1<br />
1<br />
1<br />
1<br />
4 1<br />
0<br />
1<br />
1<br />
4 1<br />
2<br />
1<br />
1<br />
1<br />
1 ,<br />
0<br />
0<br />
0<br />
2<br />
4<br />
2<br />
1<br />
6 1<br />
0<br />
1<br />
1<br />
0<br />
1 ,<br />
1<br />
2<br />
1 .<br />
1<br />
6 1<br />
1<br />
1<br />
1<br />
0<br />
1<br />
2<br />
2<br />
2<br />
1<br />
0<br />
1 ,<br />
1<br />
1<br />
1 ,<br />
1<br />
6 1<br />
1<br />
1<br />
1<br />
4 1<br />
2<br />
1<br />
1<br />
4 1<br />
0<br />
1<br />
2<br />
0<br />
2<br />
1<br />
0<br />
1 ,<br />
(2.9)<br />
Sometimes the computation cost of these latter processes is considered<br />
too high, so that approaches of compromise between speed and performances<br />
have been developed; thus for example the Deriche filter, which<br />
possesses an infinite pulse response of the shape f (x) c × exp ( |x|)<br />
permitting therefore an implementation by means of separable recursive<br />
filters. 2D filtering is obtained by the action of two filters crossed in x and<br />
y, and the filtering of the noise with the help of a function of extension<br />
(generally a Gaussian of the same standard deviation as the one considered<br />
for the detector). (See Figure 2.1.17.)<br />
0<br />
0<br />
0<br />
0<br />
0<br />
0<br />
Image improvements 99<br />
1<br />
1<br />
1 ,<br />
1<br />
2<br />
1 ,<br />
1<br />
6 1<br />
2<br />
1<br />
1<br />
2<br />
1<br />
Figure 2.1.17 From left to right: image of contours by a Laplacian, original<br />
image, and contours by the gradient of Deriche.<br />
1<br />
2<br />
1 ,
100 Alain Dupéret<br />
2.1.4 Conclusion<br />
The image improvement can often be performed by techniques of convolution<br />
with the suitable filters, of finite extension, and thus that permit an<br />
implementation on the image as linear operators. The improvement of<br />
an image aims at improving its aesthetic through subjective criteria.<br />
The improvement of the image may also use:<br />
• techniques of manipulation of colour-level levels using properties of<br />
the histogram to which the user can give the shape that he wants;<br />
• techniques of filtering in order to reduce the noise and to improve the<br />
contours in the image. The most frequently used methods are nonlinear.<br />
Bibliography<br />
Haralick R.M. (1984) <strong>Digital</strong> step edges from zero crossing of second directional<br />
derivates. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.<br />
PAMI 6, no. 1, January, pp. 58–68.<br />
Deriche R. (1987) Using Canny’s criteria to derive a recursively implemented<br />
optimal edge detector. The International Journal of Computer Vision, vol. 1,<br />
no. 2, May, pp. 167–187.<br />
2.2 COMPRESSION OF DIGITAL IMAGES<br />
Gilles Moury (CNES)<br />
2.2.1 Interest of digital image compression<br />
The digitization of pictures offers a large number of advantages:<br />
• possible processing by powerful software;<br />
• reliability of the storage (on CD-ROM, hard drives . . .);<br />
• errorless transmission (thanks to the error-correcting codes).<br />
Nevertheless, it has the inconvenience of generating often large volumes<br />
of data. To give examples, a spatial remote sensing picture acquired by<br />
the Spot satellite in panchromatic mode represents a volume of: 6,000 lines<br />
times 6,000 columns at 8 bits/pixels 288 Mbits. A classical digitized<br />
aerial image, scanned with a 14 m pixel size, provides 2,048 Mbits.<br />
Of course the first goal of the engineer, when designing the data flow<br />
system, should be to define the number of effective bits by pixel through<br />
a careful consideration of the noise in the image, so that the expected<br />
noise level is at the level of the last bit, for example. But quite often the
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Compression of digital images 101<br />
final figure is larger, in order to be able to cope with situations significantly<br />
different from the nominal one, which means very often one or two<br />
bits beyond what should be optimal.<br />
Considering limitations that apply in most systems on capacities of<br />
storage and/or transmission, it is necessary first to reduce to the minimum<br />
the quantity of necessary bits per pixel to represent the picture. To<br />
compress, one chooses the representation that provides the minimum<br />
number of bits, while preserving in the picture the necessary information<br />
for the user. The efficiency of the compression will be measured by the<br />
rate of compression, that is the ratio between the number of bits of the<br />
picture source to the number of bits of the picture compressed.<br />
The compression of pictures is possible for the following reasons:<br />
• Data pictures issued from the instrument of image acquisition present<br />
a natural redundancy that doesn’t contribute to information and that<br />
it is possible to eliminate before storage and transmission. One can<br />
compress pictures efficiently therefore without any loss of information<br />
(so-called reversible compression). Nevertheless, we will see in the<br />
following that this type of compression is limited to relatively low<br />
rates of compression (typically between 1.5 and 2.5 according to the<br />
content of the image).<br />
• Besides, the end user of pictures is interested, in general, in only part<br />
of the information carried by the picture. It is what one will call the<br />
relevant information. It will therefore be possible to compress pictures<br />
more efficiently again while removing non-relevant information (socalled<br />
compression with losses) and this to the same satisfaction of<br />
the user. Rates of compression will be able to reach much higher figures<br />
(up to 50 in certain applications of very low quality). Of course, the<br />
level of distortion due to the compression will be a function of the<br />
rate and the performance of the algorithm used.<br />
For every application, there will be a compromise to find, according to<br />
limitations of resources (storage, transmission), between the satisfaction of<br />
users (the quality of pictures being inversely proportional to the rate of<br />
compression used) and the quantity of pictures that can be stored and/or<br />
transmitted.<br />
2.2.2 Criteria of choice of a compression algorithm<br />
Criteria to take into account for the choice of an algorithm of compression<br />
and a rate of compression are very varied and depend in part on the<br />
targeted application. Among the most generic, one can mention:<br />
• The type of images: pictures of photographic type including a large<br />
number of levels of gray or colour (8 to 24 bits by pixel), or artificial
102 Gilles Moury<br />
pictures including only a few levels of gray (e.g. binary pictures of<br />
type fax). These two types of pictures require very different algorithms.<br />
For normalized algorithms, one can cite: the JPEG standard for the<br />
photographic pictures, the JBIG standard for the artificial pictures with<br />
few gray levels and the standard T4 and/or T6 for the black and white<br />
fax. We will now consider only the pictures of photographic type.<br />
• The level of quality required by the user: this ranges from a strictly<br />
reversible compression need (case of certain applications in medical<br />
imagery) to needs at a very low quality level (case of certain transmission<br />
applications of pictures on the Internet). The tolerable loss<br />
level and nature will have to be refined for example with the user<br />
through validation experiments for which several levels and types of<br />
degradation will be simulated on the representative pictures of the<br />
application. The type of degradation (artefacts) is going to be rather<br />
the function of the algorithm, and the level of deterioration the function<br />
of the compression rate.<br />
• The type of algorithm: normalized or proprietary. Principal advantages<br />
of normalized algorithms are of course compatibility and permanency,<br />
with the ascending compatibility guarantee as well. The major inconvenience<br />
is the slow evolution of norms that means that a finalized norm<br />
is very rarely the the most effective solution available.<br />
• The type of transmission or type of access to the decompressed picture:<br />
sequential or progressive. In the first case, the picture is transmitted<br />
in block to maximal resolution; in the other case, one first transmits<br />
a low-resolution version (therefore very compact) that permits the user,<br />
for example, to choose in a data base and then select a full resolution<br />
version. Algorithms using transformation (DCT, wavelets . . .) permit,<br />
among others, progressive transmissions.<br />
2.2.3 Some theoretical elements<br />
2.2.3.1 Correlation of picture data<br />
In a natural picture (as opposed to an artificial grid), the correlation<br />
between neighbouring pixels is very high and decreases according to the<br />
Euclidian distance between pixels. The correlation decreases very quickly<br />
with the distance and often becomes negligible to 5–10 pixels of distance.<br />
To compress a picture efficiently, and therefore to avoid the need to code<br />
and to transmit several times the same information, the first operation to<br />
achieve is always to locally decorrelate the picture. We will see several<br />
types of decorrelator in the following. The simplest decorrelation to achieve<br />
consists, instead of coding every pixel p(i, j) independently of its neighbours,<br />
in coding the difference: p(i, j) p(i, j1) with i an indication of<br />
line and j an indication of column.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
2.2.3.2 Notion of entropy<br />
The entropy is a measure of the quantity of information contained in a<br />
set of data. It is defined in the following way:<br />
• consider a coded picture on k bits/pixel, in which every pixel can take<br />
2 k values between 0 and 2 k 1;<br />
• the entropy of order 0 of the picture (denoted H 0(S)) is given by the<br />
formula:<br />
H0(S) <br />
2k1 Pi log2 i0<br />
1<br />
Pi , (2.10)<br />
where Pi is the probability that a pixel of the picture takes the i value.<br />
The theory of information (see reference [1]) shows that H0(S) gives, for a<br />
source without memory and therefore decorrelated, the mean minimal number<br />
of bits per pixel with which it is possible to code the picture without<br />
losing information. In other words, the maximal compression rate that it<br />
will be possible to reach in reversible compression is given by: CRmax <br />
k/H0 (S).<br />
To illustrate the notion of entropy, let’s take four different picture types:<br />
1 a uniform picture where all pixels have all the same value, H 0(S) 0,<br />
this picture contains no information;<br />
2 a binary black and white picture of the type of those processed by fax<br />
machines:<br />
H(S) P white log 2<br />
1<br />
P white<br />
Compression of digital images 103<br />
P black log 2<br />
. (2.11)<br />
One has H(S) < 1. In practice, P black
104 Gilles Moury<br />
• entropy of the picture source: H 0(S) 6.97;<br />
• entropy of the picture after decorrelation of type [p(i, j) p(i, j1)]:<br />
H 0 (S) 5.74.<br />
The more efficient the decorrelation, the weaker will be the entropy of the<br />
picture after decorrelation and the higher will be the rate of compression<br />
attainable by reversible coding of data decorrelated. Nevertheless, rates of<br />
compression attainable remain relatively low (of the order of 1.5 to 2.5<br />
on pictures of remote sensing of the type from the Spot satellite).<br />
2.2.3.4 Compression with losses<br />
When one searches for some compression rates higher than CRmax, which<br />
is the most current case, one is obliged to introduce losses of information<br />
in the chain of compression. These losses of information are achieved in<br />
general by a quantification of decorrelated data. For a given picture and<br />
algorithm of compression, there is a relation between the achieved rate of<br />
compression and the errors introduced in the picture by the compression/<br />
decompression. This curve, called ‘distortion/rate’ has the typical shape<br />
given in Figure 2.2.1.<br />
The distortion is often measured quantitatively by the standard deviation<br />
of the compression error, given by the formula:<br />
Degradation<br />
(standard deviation of the error)<br />
2.0<br />
1.5<br />
1.0<br />
0.5<br />
0<br />
1<br />
CR max<br />
2 3 4 5 6<br />
Compression rate<br />
Figure 2.2.1 Typical ‘distortion rate’ curve.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
1<br />
NM N<br />
i1<br />
M<br />
p(i, j) p′(i, j)<br />
j1<br />
2<br />
(2.12)<br />
for an image with N lines, M columns and a null mean error of compression,<br />
which is the most general case (p′(i, j) being the value of the pixel<br />
rebuilt after compression/decompression).<br />
The maximal level of tolerable distortion by the user of pictures fixes<br />
the maximal usable rate of compression on the mission. To compare performances<br />
of two algorithms of compression, one should draw the<br />
‘distortion/rate’ curves of the two algorithms and this on the same image(s),<br />
as the performance (rate) of an algorithm vary considerably from one<br />
picture to another according to the entropy of data to compress.<br />
2.2.4 Presentation of various types of algorithms<br />
Compression of digital images 105<br />
A complete presentation of the various types of compression algorithms<br />
can be found in references [1], [2] and [3]. We will give the general principles<br />
below and describe some algorithms among the more used.<br />
2.2.4.1 General architecture<br />
Any compression system can be analysed in three distinct modules (see<br />
Figure 2.2.2): the decorrelation of the source picture, the quantization of<br />
decorrelated values and the binary code assignment.<br />
1 The decorrelator allows the redundancy of data to be reduced. In practice,<br />
there are a large number of methods, more local, to decorrelate<br />
the incoming pixels. We give some examples farther on (DPCM, DCT<br />
and wavelet transform). This phase of picture processing is perfectly<br />
reversible. At the end of this first phase, the process of compression<br />
hasn’t in fact begun, but data were decorrelated so that the following<br />
processes (quantization, coding) are optimal.<br />
2 The quantizer is the essential organ of the compression system. Indeed,<br />
here the quantity of information transmitted is in fact going to decrease,<br />
while eliminating all the information not relevant (in regard to the<br />
utilization that is made of pictures) included in the data coming from<br />
the decorrelator. Let’s recall that non-relevant information is a quantity<br />
that only depends on the envisioned application(s). For example,<br />
IMAGE<br />
SOURCE<br />
DECORRELATOR<br />
DECORRELATED<br />
DATA<br />
QUANTIZATION<br />
QUANTIFIED<br />
DATA<br />
Figure 2.2.2 General diagram of a compression system.<br />
AFFECTATION<br />
OF<br />
CODES<br />
BINARY<br />
TRAIN
106 Gilles Moury<br />
in the case of the Spot satellite missions, radiometric precision is not<br />
relevant in strong transition zones (very contrasted contours). The<br />
quantity of non-pertinent information eliminated by the quantizer must<br />
be able to vary according to the application. We see therefore that the<br />
quantizer plays the role of organ of command of the compression<br />
system while allowing it to be used at different rates. The quantization<br />
is the only non-reversible operation of the compression chain.<br />
3 The assignment of codes (or more simply coding) has the role of<br />
producing a binary stream, representative of the quantized values,<br />
which will be transmitted effectively or stored for later transmission.<br />
The rate of compression really reached by the system can be valued<br />
in a realistic way only at the end of this module. Its role is to assign<br />
to every quantized value or event a binary code that will be able to<br />
be decoded without ambiguity by the decoder. This assignment can<br />
be made more or less economic in terms of number of bits transmitted.<br />
The most efficient codes are codes of variable length whose principle<br />
is simple: to assign the shortest codes to the likeliest values (or events).<br />
An example of variable-length code is the Huffman code, used in the<br />
standard JPEG and MPEG. The implementation of the various methods<br />
of compression will not show the three already mentioned modules.<br />
Some modules will be able to be regrouped in only one (as is the case<br />
for quantization and coding in algorithms using the vectorial coding).<br />
The quantization module can also disappear as in the case of a<br />
reversible compression method.<br />
2.2.4.2 Reversible algorithms<br />
There are two main classes of algorithms to achieve the reversible compression<br />
of pictures: the universal algorithms capable of compressing any type<br />
of data, and algorithms specifically optimized for pictures, these last giving<br />
performances 20 to 30 per cent better in term of compression rate.<br />
Universal algorithms<br />
The most widely used is the Lempel–Ziv algorithm (LZ, LZ77, LZW) used<br />
in the utilitarian zips, gzip, and pkzip and in formats of picture tif, gif,<br />
and png. Its principle consists in marking sequences of symbols (characters<br />
of a text, values of pixel) that repeat in the file. These sequences are<br />
then stored in a dictionary that is brought dynamically up to date. Every<br />
sequence (of variable length) is coded by a code of fixed length dependent<br />
of the size of the dictionary (e.g. 12 bits for a dictionary of 4,096 elements).<br />
This type of algorithm derives benefit from the correlation of the source<br />
and adjusts itself in real time to the local statistics (by updating the dictionary).<br />
Nevertheless, the coding of pixels being linear along the scan line,<br />
one doesn’t benefit from the vertical correlation in the picture, and then
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Compression of digital images 107<br />
the performance is not as good as that of specific algorithms that use correlation<br />
in the two dimensions.<br />
Specific algorithms<br />
The most effective algorithm is the new JPEG-LS norm [4]. The block<br />
diagram of this algorithm is given in Figure 2.2.3. It uses a 2-dimensional<br />
decorrelator of predictive type (the most often used in lossless algorithms).<br />
The value of the current pixel x is predicted from a linear combination of<br />
pixels a, b, c (see Figure 2.2.3) previously encoded. The error of prediction<br />
is then coded with the help of an adaptive Huffman code according<br />
to the context (a, b, c, d) of the current pixel (analysed by the module of<br />
context modelling). A coding by area is used in completely uniform zones.<br />
On the Spot satellite image of Genoa (Figure 2.2.6) one gets the comparison<br />
between universal algorithms and specific ones shown in Table 2.2.4.<br />
Thus we may achieve a gain of about 30 per cent on the rate.<br />
Source<br />
Image<br />
Context<br />
modelling<br />
Predictive<br />
decorrelator<br />
Coding by area<br />
c<br />
a x<br />
Huffman<br />
coding<br />
Figure 2.2.3 Block diagram of the JPEG-LS algorithm and context for the<br />
modelling and the prediction.<br />
Compressed<br />
image<br />
Table 2.2.4 Various compression rates for 2 algorithms on the Spot image of<br />
Genoa<br />
Algorithm Rate of compression<br />
JPEG-LS 1.65<br />
Lempel-Ziv (gzip-9) 1.27
108 Gilles Moury<br />
2.2.4.3 Algorithms with losses<br />
These algorithms are classified, from the simplest to the most complex,<br />
according to the three main types of decorrelators used: differential (DPCM),<br />
cosine transform (DCT), wavelet transform (multiresolution).<br />
DPCM<br />
The differential decorrelator DPCM consists in predicting the value of the<br />
pixel to code from a linear combination of values of its already coded<br />
neighbours (see Figure 2.2 3.). The prediction error is then quantized with<br />
the help of a non-uniform law adapted to the Laplace statistics of this<br />
prediction residual. This type of algorithm, used on Spot satellites 1/2/3/4<br />
to a rate of 1.33, has the feature of being very simple and to concentrate<br />
errors of compression in strong radiometric transition zones (well<br />
contrasted contours), zones in which the absolute value of the radiometry<br />
is of low importance in remote sensing. Its disadvantage is its poor performance<br />
in terms of ‘distortion/rate’ for rates ranging from middle to high<br />
(>5) (see Figure 2.2.10), owing to a too local decorrelation of the signal.<br />
For the compression with losses, the DPCM decorrelator has therefore<br />
been abandoned to the profit of the DCT.<br />
DCT<br />
The DCT (discrete cosine transform) is the decorrelator most commonly<br />
used in compression of pictures. It is the basis of many standards of<br />
compression, among which the best known are:<br />
• ISO/JPEG for still pictures ([5]);<br />
• CCITT H261 for video conferencing;<br />
• ISO/MPEG for video compression.<br />
The DCT is one of the unitary transforms (see [1] for greater precision<br />
on the definition and the interest of unitary transforms in compression)<br />
used in compression to achieve the decorrelation. This transformation operates<br />
on blocks of n × n pixels (n 8 in general) and achieves on these<br />
blocks a Fourier transform of period 2n of which only the real part is<br />
kept. The general diagram of compression based on a unitary transform<br />
DCT is given in Figure 2.2.5.<br />
The direct DCT transformation of a block 8 × 8 is given by the following<br />
formula, where p(i, j) are pixels of the source picture and the F(u, v) are<br />
DCTS coefficients representative of the spectral content of the block at the<br />
different spatial frequencies:
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Image<br />
source<br />
Rebuilt<br />
image<br />
Cut the<br />
image into<br />
blocks 8×8<br />
Concatenate<br />
8×8 blocks<br />
p′(i, j)<br />
p(i, j)<br />
F(u, v) 1<br />
2 C uC v 7<br />
i0<br />
Direct DCT<br />
Transform<br />
Inverse<br />
DCT<br />
transform<br />
F(u, v)<br />
F′(u, v)<br />
7<br />
p(i, j) cos<br />
j0<br />
Compression of digital images 109<br />
Quantization<br />
Inverse<br />
Quantization<br />
Figure 2.2.5 Chain of compression by DCT transform.<br />
(2i 1)u<br />
cos<br />
16<br />
F*(u, v)<br />
Variable flow<br />
F*(u, v)<br />
(2j 1)v<br />
,<br />
16<br />
where C u (resp. C v) = 1/√2 if u = v = 0 and C u = C v = 1 if not.<br />
In the transformed domain (spatial frequency domain):<br />
Affectation of codes<br />
(codes with variable<br />
length)<br />
Decodage<br />
(2.13)<br />
• the F(u, v) coefficients obtained are correctly decorrelated contrary to<br />
source pixels of the block;<br />
• coefficients that have a non-negligible amplitude are statistically concentrated<br />
in a restricted region of the transformed plan, which greatly<br />
facilitates the ulterior coding after quantization of these coefficients.<br />
The quantization is a simple division by a quantification step of q. The<br />
larger the q, the higher will be the rate of compression and vice versa. The<br />
rate and therefore the error on the rebuilt data is controlled therefore by<br />
the choice of the step. At the output of the quantization, a large number of<br />
coefficients (F*(u, v)) are nul. A Huffman coding is used therefore on events<br />
of type run-length adapted to this statistics. This coding, which is particularly<br />
efficient, has been optimized during the studies on the standard<br />
ISO/JPEG. This standard, of which a detailed description is given [6], is<br />
used very extensively (80 per cent of images transmitted over the Internet<br />
are compressed with JPEG).<br />
The main drawback of algorithms with a DCT basis is due to the fact<br />
that every 8 × 8 block is coded independently from its neighbours, which<br />
creates problems of block adjusting after decompression. This phenomenon<br />
(called block effect), invisible to the eye for weak compression rates<br />
(< 8), can be bothersome for some picture-processing applications. It is<br />
illustrated in Figures 2.2.6, 2.2.7, 2.2.8 and 2.2.9 that represent respectively:<br />
the source picture, the compressed DCT/JPEG to a rate of 8 picture,
110 Gilles Moury<br />
Figure 2.2.6 Excerpt of picture from Spot satellite (panchromatic mode) on the<br />
city of Genoa – non-compressed original.<br />
Figure 2.2.7 Picture of Genoa compressed with an algorithm of JPEG type to a<br />
rate of 8 (rms 6.8).<br />
a zoom on the source picture and the same zoom on the compressed picture<br />
making clear the blocks effect. To suppress this block effect and to improve<br />
again the decorrelation with regard to the DCT, one has recourse to algorithms<br />
based on the decomposition in sub-bands (by wavelet transform)<br />
described hereafter.<br />
Wavelet transform<br />
The approach of coding techniques in sub-bands is identical to that of<br />
coding by unitary transform by block (DCT or other): analysing the signal
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 2.2.8 Zoom on the non-compressed picture of Genoa.<br />
Compression of digital images 111<br />
Figure 2.2.9 Zoom on the picture of Genoa JPEG compressed to a rate of 8 (very<br />
visible 8 × 8 block effect).<br />
in different frequency components in order to code each of them separately.<br />
Means used to get this decomposition are, however, different: in the<br />
method by block a matrix 8 × 8 of transformation provides the different<br />
frequency components, whereas a real filtering of the whole picture must<br />
be achieved in the sub-bands technique. This decomposition in N subbands<br />
takes place by means of a hierarchical filtering of the picture. The<br />
filters used are determined from the theory of wavelets [7]. One thus<br />
achieves a wavelet transform of the picture.
112 Gilles Moury<br />
Advantages of decomposition algorithms in sub-bands by wavelets are<br />
as follows:<br />
• the decomposition of the picture in different spatial frequencies is made<br />
globally on the picture (by a mobile window of filtering) and not by<br />
8 × 8 block as in the DCT. Thus there are no problems of continuity<br />
in borders of blocks on a decompressed picture;<br />
• ‘distortion/rate’ performances are much better for the rates superior<br />
to 8 (typically): improvement of 30 per cent to 50 per cent of the<br />
compression rate for the same distortion level (see Figure 2.2.10);<br />
• the progressive transmission of different resolution levels and the noise<br />
removal is very easily integrable in this type of compression scheme.<br />
For these different reasons, the JPEG committee has been reactivated to<br />
define the JPEG2000 standard based on wavelet transform. This future<br />
ISO standard should be completed in 2000 [8] and integrates a large<br />
number of functions (random access to the compressed file, progressive<br />
transmission, resilience to transmission errors, ascending compatibility with<br />
JPEG . . .).<br />
For these algorithms based on wavelet transform, the most effective<br />
quantization/coding techniques are bit planes quantization and zero-trees<br />
coding. The state-of-the-art representative algorithms are EZWS [9] and<br />
SPIHT [10].<br />
Comparison<br />
One compares ‘distortion/rate’ performances of the three types of decorrelators<br />
(DPCM, DCT, wavelets) through three representative algorithms,<br />
on the Spot satellite picture of Genoa (Figure 2.2.6) (see Figure 2.2.10):<br />
• DPCM: JPEG-LS algorithm in quasi-lossless mode [4];<br />
• DCT: JPEG algorithm ‘baseline’ mode [5];<br />
• wavelet: SPIHT algorithm [10].<br />
2.2.5 Multispectral compression<br />
The algorithms presented until now only take benefit from the intra-picture<br />
correlation (or spatial correlation). In the case of a multispectral (or colour)<br />
picture, it is also possible to exploit the correlation between the different<br />
spectral bands. In the case of the Spot satellite pictures for example, which<br />
include four bands (b1, b2, b3, mir), b1, b2 bands are very strongly correlated<br />
on certain types of landscapes whose spectral content has only<br />
moderate variations.<br />
It is therefore interesting to achieve a spectral decorrelation then a spatial<br />
decorrelation of the picture (the inverse order is also possible). A complete
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
rms of the error<br />
20<br />
16<br />
12<br />
8<br />
4<br />
0<br />
0 4 8<br />
compression rate<br />
12<br />
analysis of all usable techniques can be found in [11]. The spectral decorrelator<br />
can be of predictive type or by orthogonal transform. The optimal<br />
transform to achieve this operation is the Karhunen–Loève transform (KLT).<br />
This transform enables one to switch from a spectral coordinates system<br />
[bi] where components are greatly correlated, to a reference [di] where<br />
components are decorrelated and the information (energy) of the picture is<br />
concentrated on a few components. An example of such a transformation<br />
is that used in digital television (standard CCIR 601 [12]): transfer from<br />
the system [R,V,B] to the system [Y,Cb,Cr], 80 per cent of information<br />
is concentrated on the signal of Y luminance; signals of chrominance<br />
differences (Cb,Cr) carry little information and thus can from then on be<br />
under-sampled and/or compressed to higher rates than the luminance.<br />
2.2.6 Perspectives<br />
Compression of digital images 113<br />
DPCM<br />
DCT<br />
Wavelets<br />
Figure 2.2.10 Comparative curves ‘distortion/rate’ for DPCM, DCT, wavelets.<br />
Beyond the compression by wavelet transform now extensively used and<br />
which will be soon normalized (JPEG2000), different solutions are currently
114 Gilles Moury<br />
underway to attempt to improve the ‘distortion/rate’ performances of algorithms.<br />
Among these, one can cite:<br />
• the compression by fractals [13] that, coupled to the decomposition<br />
in sub-bands by wavelet, permits progress in the high rate range (> 30);<br />
• the selective compression that consists in detecting in the picture, zones<br />
of interest for the user, in order to apply to those a rate of compression<br />
preserving the quality of the picture, while the remainder of the<br />
picture can be compressed with a higher rate (e.g. on the Spot satellite<br />
pictures, zones of interest are zones without clouds). The selective<br />
compression opens the way to the more elevated rate used in many<br />
applications. All the difficulty lies in the reliability of the detection<br />
module.<br />
References<br />
[1] M. Rabbani, P.W. Jones, <strong>Digital</strong> Picture Compression Techniques, SPIE Press,<br />
vol. TT 7, 1991.<br />
[2] J.P. Guillois, Techniques de compression des images, ed. Hermès, coll.<br />
Informatique, 1995.<br />
[3] K.R. Rao, J.J. Hwang, Techniques and Standards for Image, Video and Audio<br />
Coding, Prentice Hall, 1996.<br />
[4] International Standard ISO-14495–1 (JPEG-LS).<br />
[5] International Standard ISO-10918 (JPEG).<br />
[6] W.B. Pennebaker, J.L. Mitchell, Still Image Data Compression Standard, Van<br />
Nostrand Reinhold.<br />
[7] M. Vetterli, J. Kovacevic, Wavelets and Subband Coding, Prentice Hall, 1995.<br />
[8] http://www/jpeg.org (official site JPEG).<br />
[9] J.M. Shapiro, Embedded image coding using zerotrees of wavelet coefficients<br />
(EZW), IEEE Transactions on Signal Processing (Special Issue on Wavelets<br />
and Signal Processing), vol. 41, December 1993.<br />
[10] A. Said, W.A. Pearlman, A new fast and efficient image codec based on Set<br />
Partitioning in Hierarchical Trees (SPIHT), IEEE Transactions on Circuits and<br />
Systems for Video Technology, vol. 6, June 1996.<br />
[11] J.A. Saghri, A.G. Tescher, J.T. Reagan, Practical transform coding of multispectral<br />
imagery, IEEE Signal Processing Magazine, vol. 12, January 1995.<br />
[12] ITU-T Recommendation 601, Encoding parameters of digital television for<br />
studios, 1982.<br />
[13] M.F. Barnsley, L.P. Hurd, Fractal Image Compression, Peters, 1993.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
2.3 USE OF GPS IN PHOTOGRAMMETRY<br />
Thierry Duquesnoy, Yves Egels, Michel Kasser<br />
2.3.1 Use of GPS in photographic planes<br />
Thierry Duquesnoy<br />
Use of GPS in photogrammetry 115<br />
2.3.1.1 Introduction<br />
In aerial images acquisition, the GPS (global positioning system), or any<br />
other GNSS equivalent system (GLONASS, or the future GALILEO) may<br />
be used in three distinct phases:<br />
• the real-time navigation;<br />
• the determination of the trajectory of the plane, more precisely the<br />
determination of the position of the coordinates of the camera at the<br />
time of the image acquisition;<br />
• the determination of the attitude of the plane.<br />
If these three aspects use the GPS effectively, they do not require the same<br />
type of measure. The most original use concerns the measure of attitude<br />
that we will describe farther.<br />
We will recall briefly here that two types of GPS observations exist:<br />
• the pseudo-distances that are measures of time of propagation of the<br />
pseudo-random modulation of the signal;<br />
• the measure of phase, performed on the wave carriers, and that is<br />
performed on an identical but stable signal generated by the receptor<br />
and locked on the satellite signal.<br />
And two types of positioning exist, the absolute positioning and the relative<br />
positioning. The precision of the positioning varies according to the<br />
type of observations, from a positioning accurate at the level of a few<br />
metres to a relative precision of the order of 2 mm ± 10 6 D to 10 8 D<br />
using the measure of phase.<br />
In the end of section (2.3.1.5), one will find a summary of the terminology<br />
employed for the different uses of the GPS.<br />
2.3.1.2 The navigation<br />
The navigation is a priori the simplest mode. It uses the pseudo-distances.<br />
At the time of the image acquisition, it is necessary to have the position,<br />
in almost real time, of perspective centres in order to verify that the followup<br />
of the axis by the plane is good and that the overlap rate is correct.<br />
In absolute positioning, with a precision that is about ten metres (with the<br />
SA switched off, as it is since May 2000), conditions are achieved. The<br />
metric precision realized is by far superior to the 40 m with which pilots
116 Thierry Duquesnoy et al.<br />
can hold the axis of flight. If a better precision is required, the differential<br />
mode is necessary: several solutions may exist to get the differential corrections<br />
in real time. One can use networks of stations that broadcast<br />
corrections on different frequencies. One can arrange his own GPS receptor<br />
on a known fixed station that sends corrections. And then there are differential<br />
correction services via a geostationary satellite. This last process<br />
seems best suited for airborne applications.<br />
Let’s note that in differential navigation, the speed of the plane is known<br />
to 0.1 knot, which is sufficient for photogrammetric applications.<br />
For the overlap, it may be necessary to know the attitude of the plane<br />
at the same time. We will describe these needs in the chapter dedicated to<br />
the determination of attitude.<br />
2.3.1.3 The determination of perspective centres for the<br />
aerotriangulation<br />
We saw that measures of pseudo-distances permit a metric positioning<br />
when differential data are available. However, the necessary precision on<br />
the determination of perspective centres must be at the same level as that<br />
necessary for the photogrammetric process, typically 10 m × the scale of<br />
the picture. A decimetre precision is therefore necessary to most photogrammetric<br />
applications (scale varying from 1:5,000 to 1:40,000). This precision<br />
cannot be reached with the exclusive utilization of pseudo-distances. But<br />
the measure of phase is ambiguous as, while using the phase, one only<br />
measures the fractional part of the cycle, the whole number of cycles separating<br />
the receptor of the satellite being unknown.<br />
Therefore there are two ways to improve the precision of the determination<br />
of perspective centres a priori: improve the process of the pseudodistances<br />
while using the phase, or be able to solve ambiguities in the<br />
process of the phase.<br />
The trajectography<br />
This is a calculation in differential mode between a reference station of perfectly<br />
known coordinates and the antenna and receptor on board the plane.<br />
The process is based on the fact that phase data are less noisy than pseudodistance<br />
data. Indeed, one considers that the noise of the measure is estimated<br />
to be better than 0.1 per cent of the wavelength of the signal on which the<br />
measure is made. Therefore, the noise on the pseudo-distances is metric while<br />
the noise on the phase is millimetric. Thus one tries to improve the process<br />
of pseudo-distances by decreasing the noise of measure, by a smoothing<br />
process with the help of the phase measurement. This is the trajectographic<br />
mode. In this mode no initialization is necessary. One gets by this method<br />
an internal consistency within one axis inferior to the decimetre. But then,<br />
the precision between neighbouring strips is no better than some metres.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
The resolution of ambiguities on the flight<br />
Again, this is a positioning in differential mode. In the case of two fixed<br />
receptors, the resolution of ambiguity is possible without too many difficulties<br />
for baselines of several hundred kilometres. But it is much more<br />
difficult when one of the GPS receptors is mobile.<br />
An algorithm developed by Remondi (1991), permits the resolutions of<br />
ambiguities in flight (on the fly, or OTF) without any initialization in static<br />
phase, which would obviously not be possible.<br />
The initialization of the ambiguities fixation in flight cannot be performed<br />
beyond a distance of 10 km between the fixed receptor and the mobile.<br />
Once the initialization is made, the basis can go until 30 km, but without<br />
cycle slip. If there is a cycle slip, an OTF traditional software will do the<br />
initialization again, and therefore the mobile receptor should come once<br />
more closer than 10 km from the reference. This method of ambiguity<br />
resolution in flight is widely used by numerous constructors and GPS users,<br />
but it has the major inconvenience, in an application of photogrammetric<br />
production, of imposing maximal lengths of bases of about 30 km. Some<br />
constructors provide solutions for bases ranging to 50 km, but no full-size<br />
test has been achieved again. Nevertheless, the order of magnitude remains<br />
the same and requires the presence of a nearby fixed receptor on the zone<br />
to photograph.<br />
The GPS positions, calculated during the mission are those of the centre<br />
of phase of the antenna. However, the position that interests us is that of<br />
the perspective centre at the time of the image acquisition. It requires a<br />
good synchronization between the image acquisition and the GPS receiver,<br />
and a perfect knowledge of the vector antenna/camera at the time of the<br />
image acquisition. This vector is constant in the airplane frame, on the<br />
other hand it is not in the earth reference system. It is therefore necessary<br />
to calculate the matrix of passage between the two systems, either using<br />
ground points, or knowing the attitude of the plane at the time of the<br />
image acquisition.<br />
2.3.1.4 Measures of airplane attitude<br />
Use of GPS in photogrammetry 117<br />
The measure of attitude by GPS is an interferometric mode. One measures<br />
times of arrival of the wave on a system of several antennas forming a<br />
strong solid shape. It requires a perfect synchronization on times of arrival<br />
to the different antennas. It is achieved with the use of receptors, a clock,<br />
and several antennas. One finds currently on the market systems of three<br />
or four antennas.<br />
The more the antennas are separated from one another, the better is the<br />
angular precision. Unfortunately, it is very difficult to get solid polygons on<br />
planes used for aerial pictures, due to the flexibility of the wings (with the<br />
notable exception of airplanes with high wings). The currently achievable
118 Thierry Duquesnoy et al.<br />
precisions are in the order of the thousandth of a radian whereas a precision<br />
ten to a hundred times superior would be necessary.<br />
It is possible, but at a considerable cost, to couple to the GPS an inertial<br />
system (IMU, for inertial measurement unit). The utilization of the<br />
inertial system alone is not possible because it drifts very quickly. But then,<br />
coupling the inertial platform with the GPS measures will cancel most of<br />
the drifts, and makes it possible to get a precision on the angular measures<br />
of about 10 5 radians. This is the device used in aerial laser scanning.<br />
Information of attitude is known in real time in the plane and can therefore<br />
also be used for the navigation in order to minimize overlap errors,<br />
and to have a sufficiently precise assembly picture at the end of the aerial<br />
mission.<br />
2.3.1.5 GPS terminology<br />
• Kinematics: differential method (DGPS) based on the measure of phase<br />
of at least four satellites. The principle consists in solving the ambiguities<br />
by an initialization, then to observe the points for a few seconds<br />
while preserving the signal on satellites during the journeys, and therefore<br />
the same integer ambiguities.<br />
• Pseudo-kinematics: original method of kinematic type that consists of<br />
reobserving the reference points and treating the file of observation as<br />
if it was a fixed station having a gap in its data (ambiguities being<br />
the same).<br />
• Dynamic: differential method (DGPS) with observations of pseudodistances,<br />
the station of reference whose coordinates are known, sends<br />
by radio in real time the corrections to the mobile station, which may<br />
then calculate its position.<br />
• Trajectography: specific mode that uses observations of phase and<br />
pseudo-distances, the computation being performed using pseudodistances<br />
smoothed by the phase.<br />
• Navigation: we will associate to this term the common sense of navigation,<br />
that is the knowledge of information on the position and the<br />
attitude of the plane that could allow the pilot as well as operators<br />
to rectify their trajectory or the orientation of the camera in situations<br />
requiring it.<br />
2.3.2 Use of GPS in aerotriangulation<br />
Yves Egels, Michel Kasser<br />
2.3.2.1 How to limit the stereopreparation work<br />
Whatever the photogrammetric product to be achieved (vectorial restitution,<br />
orthophotography, DTM), it is indispensable to determine in a previous
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Use of GPS in photogrammetry 119<br />
phase the formulas describing the geometry of the used pictures. This phase<br />
is traditionally called aerotriangulation (or spatiotriangulation in the case<br />
of spatial images, whose geometry is often different). We will recall quickly<br />
here the principle of this classic method in analytic photogrammetry (one<br />
will find the complete survey of this problem in any good book on analytic<br />
photogrammetry), and we will develop only the improvements made<br />
possible by the appearance of digital photogrammetry.<br />
Let’s recall that the geometric properties of images alone do not permit<br />
one to know either the position, or the orientation, or the scale of the<br />
photographed objects. The determination of the corresponding mathematical<br />
similitude cannot be obtained except by external measures: thus,<br />
if one limits oneself to the case of the stereopair, and to determinations<br />
of ground control points, it will be necessary to measure at least two points<br />
in planimetry and three points in altimetry.<br />
The aim of aerotriangulation is to reduce as much as possible this requirement<br />
of field measures, by processing simultaneously the geometry of a<br />
large number of images. The goal is to adjust (often to the least square<br />
sense) the measures of coordinates in the images of the homologous tie<br />
points, a certain number of ground points being considered as known, and<br />
possibly auxiliary measures recorded during the photographic flight as well,<br />
or satellite trajectography data. This leads generally to an overabundant<br />
system of several thousands of non-linear equations, that one will be able<br />
to solve by successive approximations from an approached solution.<br />
Many software packages on the market allow one to perform this calculation.<br />
Without any in-flight measures, they allow the necessary ground<br />
points to be limited to one planimetric point for five to six couples, and<br />
one altimetric point by couple. The digital photogrammetry did not bring<br />
any basic difference at the level of the aerotriangulation calculation itself,<br />
but on the other hand it allows the measure of tie points, which remained<br />
manual (and very trying) in analytic plotting, to be automated almost<br />
completely. Besides, the development of digital cameras, as they require<br />
more images to cover a given zone, has led to the search for more efficient<br />
auxiliary measures in order to restrict again the ground measures.<br />
These normal mean values (one point in planimetry for five or six couples,<br />
and one point in altimetry) depend in fact on the geometry of the work,<br />
and in particular on overlaps from image to image (the minimal value is<br />
60 per cent), as well as overlaps between neighbouring strips that have a<br />
major impact in the rigidity brought to the general figure. If the overlap<br />
between strips is significant (e.g. 60 per cent), every point is found in three<br />
strips, but at the end it doesn’t permit an appreciable decrease in the<br />
number of control points below half of values for a normal aerotriangulation.<br />
On the other hand, one gets a much better reliability, and the<br />
precision is improved as well.<br />
This is a solution used for the survey of points that must be measured<br />
with a very high precision, materialized by targets so that the pointing is
120 Thierry Duquesnoy et al.<br />
almost ideal (e.g. measure of ground deformations). On the other hand,<br />
the additional costs of this solution makes it a non-profitable one for classical<br />
cartographic studies.<br />
2.3.2.2 Use of complementary measures during the flight<br />
One formerly used measures of APR (airborne profiles recorder, see §1.7),<br />
a method that provided some vertical distances (radar or laser) from the<br />
plane to the ground, with as altimetric reference an arbitrary isobar surface.<br />
The airplane followed trajectories forming a grid, the points of impact<br />
of the ranging unit to the ground being controlled by pictures acquired<br />
by a camera. This methodology fell into obsolescence, but proved to be<br />
quite useful to cover vast uninhabited zones (Australian desert, northern<br />
Canada).<br />
Use of complementary measures during photographic flights<br />
It would ideally be necessary to know at each instant where a picture is<br />
acquired:<br />
• the position of the centre of the perspective, that is to say the optic<br />
centre of the camera, in the reference frame of the work;<br />
• the orientation of the camera.<br />
These two groups of parameters are not at all equivalent in terms of necessary<br />
acquisition expense.<br />
For the position of the centre, let us note that almost no one uses the<br />
differential barometer, which could bring some very interesting elements<br />
while creating some strong new constraints on what is one of the major<br />
weaknesses of aerotriangulation, that is to say the altimetry. This equipment<br />
is otherwise functional without any external infrastructure, unlike<br />
the GPS that we are going to examine, which is a more complete solution<br />
on the other hand.<br />
For the orientation of the camera, there does not exist a reliable and<br />
low-priced sensor. The available inertial platforms are costly. Based on the<br />
use of accelerometers whose values are integrated twice to get the displacements,<br />
and on gyrometers or gyroscopes to measure the variations of<br />
attitude, they have by their very nature a random drift more or less proportional<br />
to the square of the time. This requires frequent updates, and this<br />
updating can then be performed by the GPS. There are also (see §2.3.1)<br />
some solutions using the GPS alone, measuring the GPS data simultaneously<br />
acquired on three or four antennas. This is a mode that works like<br />
the interferometry. Its angular sensitivity depends in fact on the antenna<br />
spacing, the values of measure of the vertical component on each antenna<br />
being hardly better than 1 cm. This implies, as the antennas cannot be
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Use of GPS in photogrammetry 121<br />
more distant than 10 m (antennas on wings, the nose and the vertical<br />
stabilizer of the plane) a precision of measure that is hardly better that<br />
some millirad, with all the risks of GPS measures. By comparison, with a<br />
pointing accuracy on image around 15 m (which is not excellent) and a<br />
classic focal distance of 150 mm, one understands that the need in absolute<br />
orientation is between 10 4 and 10 5 rad, which disqualifies de facto such<br />
a use of the GPS. Otherwise, it would be necessary that the four antennas<br />
form a rigid polyhedron, and it is not the case since most planes’ wings<br />
bend according to constraints brought about by atmospheric turbulence:<br />
the only noticeable exception is formed by planes with high wings, when<br />
the implantation of antennas on wings is localized on the zone of shroud<br />
grappling that holds the wings.<br />
2.3.2.3 Positioning by GPS<br />
The precision that it is necessary to achieve to localize the perspective<br />
centres must be at least as good as that of points to localize on the ground,<br />
and obviously depends on parameters of the image acquisition (scale, height<br />
of flight, pixel size, etc.). In current cases it means a precision ranging<br />
from 5 to 30 cm. We are in a dynamic regime, with a speed of the plane<br />
around 100 to 200 m/s. It implies a very good synchronization between<br />
the camera and the GPS receiver, which generally does not pose any particular<br />
problem with the GPS equipment. The GPS should in this case have<br />
a special input, the external events datation being performed with a precision<br />
that reaches the 100 s level (GPS easily performs 100 times better).<br />
But it is necessary that the camera provides a signal perfectly synchronous<br />
(better than 0.1 ms) with the opening of the shutter, which is the case<br />
with all modern cameras but is not always true of old equipment.<br />
In addition it is necessary to get a continual positioning during the movement.<br />
With GPS measures at each second, one has to interpolate the<br />
position of the camera at the time of the signal of synchronization on<br />
distances of 100 to 200 m, which does not permit a decimetric precision<br />
as soon as the atmosphere is a little turbulent and creates perturbations<br />
on the line of flight. And one should note that the higher the requested<br />
precision, the more one flies at low altitudes and therefore the stronger<br />
the turbulence. Thus a positioning adapted to large scales requires a better<br />
temporal sampling than 1 s, a value of 0.1 s (yet rarely available on the<br />
GPS receivers) being barely sufficient. In this optics, one can associate a<br />
GPS receiver and an inertial platform that serve as a precise interpolator,<br />
the bias of the inertial unit being small on one second, so that the unit<br />
can provide data quite often (typically from 100 to 1000 Hz). But this<br />
solution is quite expensive and still not very often used.<br />
As we saw in §2.3.1, the precision requirements need the use of the GPS<br />
in differential mode. Two modes can be used:
122 Thierry Duquesnoy et al.<br />
• Kinematic mode, with measure of the phase on the signal carrier. Either<br />
one initializes with a time duration permitting the ambiguity resolution<br />
(it requires some minutes) and fixes them then for the remainder of<br />
the flight. But then the reliability is low, and this is not used because<br />
when the plane turns it frequently occurs that there are short<br />
interruptions of the signal, ruptures in the continuity of the measure<br />
that would oblige us to solve the ambiguities once more, which becomes<br />
obviously impossible in flight. Either one solves ambiguities on the<br />
flight (OTF, AROF, . . . methods), which overcomes this difficulty.<br />
Nevertheless, it is necessary to note that up to now these methods work<br />
only with reasonably short bases (20 or 30 km to the maximum), which<br />
is poorly compatible with work conditions of photographic planes.<br />
Indeed it would be necessary, before the beginning of the flight, to put<br />
in operation a GPS station close to the zone to photograph. Risks of<br />
photographic flight programming in the mid-latitude countries makes<br />
such a step quite restrictive, the weather alone being a sufficiently strong<br />
constraint not to add an obligation to set up such an equipment, which<br />
would lead to too much time wasted. The only cases where one can<br />
work according to this mode will be those where a permanent GPS station<br />
close to the zone to photograph exist. This method is therefore not<br />
very often used.<br />
• Trajectography mode. One then measures the pseudo-distances as well<br />
as the carrier phase. The precision is not so good, but one can work<br />
with much more important distances to the reference station. We are<br />
going to study this solution in more detail.<br />
2.3.2.4 Survey of the trajectography mode<br />
The satellite-receptor distance can be calculated on the one hand from the<br />
time of flight of the wave deduced from the datation of the C/A code, the<br />
clock correction, the tropospheric and ionospheric corrections (typical<br />
precision of the order of one metre in differential mode). It can be calculated<br />
on the other hand from the phase measurement, with an integer k<br />
number of phase turns, from the correction of a clock, and from tropospheric<br />
and ionospheric corrections (the precision is then millimetric, except<br />
that k is unknown). The difference between these two variants of the same<br />
distance being necessarily null in theory, while integrating measures on a<br />
given satellite during enough time one can get an evaluation of k. This<br />
mode is not adapted to real-time operations, on the other hand as long<br />
as there are satellites in common visibility between the reference station<br />
and the airplane, the result is usable as far as the GDOP (geometric dilution<br />
of precision, a parameter giving an evaluation of the geometric quality<br />
of the observable satellites, the smaller figures meaning the better geometry)<br />
is correct, and the observed bases can range beyond a thousand<br />
kilometres.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Use of GPS in photogrammetry 123<br />
Results are in general always of the same type: for a few minutes the<br />
noise of determination of airplane coordinates is lower than 5 cm, and<br />
one observes a bias, positioning errors varying slowly, with a typical time<br />
constant of 1 or 2 hours, the amplitude of the errors being of the order<br />
of one metre. This is due to the slow changes of satellite configuration in<br />
the sky, the orbit errors being directly input in the calculations.<br />
We may then model the GPS measures as an error of general translation<br />
of the GPS positions for every applicable short temporal period,<br />
typically of 10 min, for example for every set of photographs taken along<br />
a given flight axis. One adds thus a supplementary unknown triplet in the<br />
aerotriangulation for every strip. If one works in differential mode, the<br />
precision is then centimetric, and in absolute mode this is without significance<br />
since there are reference points on the ground that completely impose<br />
the wanted translation for the whole set of images.<br />
One thus succeeds in decreasing considerably the number of ground<br />
points: in planimetry, it is necessary to get now at least 4 points by block,<br />
and some more but only for controls, and in altimetry 1 point all 20 or<br />
30 couples is satisfactory. These points don’t necessarily need to be to the<br />
extreme borders of the zone.<br />
One should be very attentive to problems that are linked to reference<br />
frames. In particular in altimetry one should remember that the GPS<br />
measures are merely geometric, whereas the levelling of reference is always<br />
based on geopotential measures. Geoid slopes, or rather slopes of the null<br />
altitude surface, must be taken into account according to the requested<br />
precision.<br />
One can consider treating the work without any ground references at<br />
all (inaccessible zones for example). In this case one gets a coherent survey<br />
but with a translation error of metric size, translation that is the mean<br />
value of the systematism unknowns.<br />
One should be also attentive to the geometry of the set of stereopreparation<br />
points. If they are all reasonably aligned, the transverse tilt will be<br />
undetermined. It is necessary to include points therefore in the lateral<br />
overlap zones between strips. If one works on a zone without sufficient<br />
equipment in levelling, one will be able to improve the quality of the aerotriangulation<br />
by crossing the parallel strips by transverse photo strips. But<br />
it will be necessary to perform levelling traverses on the ground if the<br />
geoid slope is unknown, just to impose its slope in the geometric model.<br />
2.3.2.5 Adjusting the link between the GPS antenna to the optic<br />
centre of the camera<br />
One will start with measuring with classical topometric methods the link<br />
vector E ranging from the antenna to the optic centre, in the airplane reference<br />
system. XA being the position of the antenna, XS that of the optic<br />
centre, one will write:
124 Thierry Duquesnoy et al.<br />
X S X A T R·E, (2.14)<br />
where T is the vector describing the residual systematism of these measures,<br />
and R describes the rotation of the airplane in the reference system of the<br />
study, assumed equal to the rotation of the camera, itself even deduced<br />
from the rotation of the bundle, a by-product of the collinearity equation.<br />
2.3.2.6 Conclusion<br />
The use of the GPS (or any other GNSS, like the Russian GlONASS or the<br />
future European GALILEO) in the plane is a very important auxiliary source<br />
of data to reduce the costs of the control points on the ground for the aerotriangulation.<br />
But in such a case it is necessary to be aware of the imperfections<br />
of the GPS, and to know what to do when some subsets of data<br />
are not exploitable. On the other hand, it must be well understood that if<br />
some inertial measurements may help considerably to interpolate within the<br />
GPS data, they cannot correct its possible defects, and thus such data should<br />
be considered as additional, every possible effort should be made to have<br />
the best possible signal in the antenna: any device allowing the absolute<br />
orientation of the camera to be provided will be welcome, but while waiting<br />
for its availability, the GPS already gives excellent results . . .<br />
Reference<br />
Remondi B.W. (1991) Kinematicc GPS results without static initialization. NOAA<br />
Technical Memorandum NO S NGS-55.<br />
2.4 AUTOMATIZATION OF AEROTRIANGULATION<br />
Franck Jung, Frank Fuchs, Didier Boldo<br />
2.4.1 Introduction<br />
2.4.1.1 Presentation<br />
This section concerns the automatic determination of aerotriangulation. This<br />
automation aims at the production of tie points and their use in two domains:<br />
the aerotriangulation itself and the automatic realization of index maps,<br />
which is a preliminary step to any aerotriangulation process. The reader will<br />
note that we do not consider here the measure of the reference points. One<br />
will not treat their determination (which is made by techniques of geodesy),<br />
or the measure of their position in images (which are measured by hand).<br />
The automatic calculation of tie points is a topic of growing interest<br />
notably by reason of the increasing number of images produced for photo-
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Automatization of aerotriangulation 125<br />
grammetric studies. In particular, both the use of digital cameras that is<br />
now increasing, and of strong overlaps (notably interstrips) on the other<br />
hand, generate images in greater number than in the past. It is henceforth<br />
possible to find a substantial number of manufacturers of aerotriangulation<br />
software proposing a module of automatic determination of tie points.<br />
These modules are able to give good results on simple scenes (including<br />
neither important relief nor too large-textured zones). However, the determination<br />
of tie points in not so simple configurations still mobilizes greatly<br />
the community of researchers. These less simple configurations are notably<br />
terrestrial photogrammetric studies, aerial images of different dates, scenes<br />
with strong relief, etc.<br />
This type of problem is a topic of a working group on behalf of the<br />
OEEPE (Organisation Européenne d’Études Photogrammetriques Expérimentales)<br />
(Heipke, 1999).<br />
2.4.1.2 Basic notions<br />
We should recall some notions that will be used in what follows:<br />
• Tie point: 3D coordinates often corresponding to the position of a<br />
physical detail of the scene, and seen in at least two images.<br />
• Measure: 2D coordinates of the projection of a point (of reference, or<br />
of link) in an image.<br />
• Point of interest: a point of the image around which the signal has<br />
specific characteristics, such as high values of the derivatives in several<br />
directions or at least two orthogonal directions, and detected by a<br />
particular tool (e.g. detection of corners, of junction).<br />
• Similarity measure: function associating to two neighbourhoods of two<br />
points a real finite number. The more often used measure is the linear<br />
correlation coefficient.<br />
• Homologous points: set of points satisfying some properties in regard<br />
to the similarity measure. Generally, a point P 1 of an image I 1 and a<br />
point P 2 of an image I 2 are judged homologous if, for any P of I, P 1<br />
and P 2 are more alike than P and P 2 , and for any Q of I 2 , P 1 and P 2<br />
are more alike than P 1 and Q.<br />
• Repeatability: quality of a detector of points of interest capable, for<br />
a given scene, of detecting points of interest corresponding to the same<br />
details in the different images, even if conditions of image acquisition<br />
vary (lighting, point of view, scale, rotation . . .). A detector of points<br />
of interest is more repeatable if it produces the same sets of points for<br />
a given scene in spite of variations of the condition of image acquisition.<br />
This notion is specified in Schmid (1996).<br />
• Disparity: for two images possessing a weak rotation angle one in<br />
relation to the other, the disparity of two points is the vector equal<br />
to the difference of their coordinates. In the case of aerial image
126 Franck Jung et al.<br />
acquisitions with vertical axis, the disparity of two points corresponding<br />
to the same physical detail of the land is bound directly to<br />
the altitude of this detail. This notion may be extended to the case of<br />
images possessing any relative rotation.<br />
2.4.2 Objectives of the selection of points of interest<br />
This section discusses three points concerning objectives of methods of<br />
automatic measurement of tie points. We will treat first their reliability,<br />
then their precision, as well as a third important point, which makes an<br />
important difference between the automatic measurement of tie points and<br />
the manual techniques: it is about the number of tie points.<br />
2.4.2.1 Reliability<br />
Most methods of aerotriangulation depend on an optimization using a<br />
least square adjustment technique. This technique is especially sensitive to<br />
aberrant values. As usual it is therefore necessary, during the calculation<br />
of tie points, to try not to provide any aberrant measures.<br />
The first major objective toward a method of automatic measurement<br />
of tie points is thus to provide points that are exempt from mistakes.<br />
2.4.2.2 Precision<br />
This second major objective can be the carrying out of complementary<br />
studies as soon as one knows that the aerotriangulation work has been<br />
performed correctly (i.e. there are no more mistakes).<br />
To get a better precision for the aerotriangulation implies more precise<br />
measures of tie points, but also taking into account in methods of aerotriangulation<br />
an error model adapted to the real errors made by measuring<br />
tools (which implies a study of these errors to model them correctly).<br />
Indeed, error models used are generally Gaussian, and it remains to prove<br />
that this model is suitable for a particular measuring tool. If necessary an<br />
adaptation of the error model may be advisable.<br />
2.4.2.3 Number of tie points<br />
Traditional aerotriangulation uses 15 measures by image, whereas tools of<br />
automatic measure of tie points are able to provide a large amount of data.<br />
In the manual case, these points are very satisfying, because they are<br />
selected by an operator. The operator can notably a priori assure their<br />
geometric distribution. In the automatic case, the lack of intelligence of<br />
any computer method may be partly compensated by the abundance of<br />
data. Concerning the distribution of points, it is difficult to force the<br />
machine to ‘find’ a solution in a given zone, and it is therefore probably
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
preferable to let it find another solution in the neighbourhood of the desired<br />
area. Concerning errors, the abundance of data does not reduce the error<br />
rates of the methods, but permits the use of statistical tools aiming to eliminate<br />
these errors.<br />
Studies are not however advanced enough to permit one to fix an ideal<br />
value of measures by image. Nevertheless, one expects that the number of<br />
required points be appreciably higher in the automatic case (typically 100<br />
measures by image (Heipke, 1998)). The debate on this question remains<br />
very open.<br />
2.4.3 Methods<br />
Automatization of aerotriangulation 127<br />
The automatic measurement of tie points generally requires two major<br />
steps. The first concerns the detection and the localization of points of<br />
interest in images. Points of interest can be calculated individually from<br />
the images. Then these points are used to produce the actual tie points.<br />
The passage of points of interest to tie points can be generally performed<br />
with two opposite strategies.<br />
In the two cases, strategies search between images for points whose value<br />
of similarity measure used is maximal. Strategies differ according to the<br />
zone in which one looks for the maximum.<br />
For a point of interest P of an image I, the first strategy consists in<br />
searching in an image J for the position of the point Q that maximizes<br />
the value of similarity among all possible positions (all pixels of a region<br />
where one expects to find Q). One can even consider calculating the position<br />
of Q with a sub-pixel precision. This strategy, which one will call<br />
‘radiometric’ is oriented: I does not play the same role as J in this case.<br />
The second strategy consists in restricting the research space in J to only<br />
the points of interest calculated. One will call this approach ‘geometric’<br />
because it is focused on the position of tie points.<br />
2.4.3.1 Methods of detection of points<br />
Detectors of points use particular features of the signal. We will more<br />
especially mention the detector of Förstner (Förstner and Gülch, 1987)<br />
(much used in commercial software) as well as the detector of Harris whose<br />
repeatability is high (Harris and Stephens, 1988; Schmid, 1996).<br />
These detectors generally have a poor localization precision. Experiences<br />
have shown a shift (rms) of 1 to 1.5 pixel between the real corners and<br />
the detected points (Flandin, 1999). Nevertheless, the delocalization of<br />
these points possesses a very strong systematism but a low standard<br />
deviation (around 1 ⁄2 pixel). Thus it is possible to consider these points<br />
as satisfactory for a photogrammetric set-up. A better localization of<br />
these detected points can be considered by the use of specific techniques<br />
(see §2.4.2.2).
128 Franck Jung et al.<br />
Figure 2.4.1 Two 300 × 200 images and their points of interest.<br />
Figure 2.4.1 represents two neighbouring images (size 300 × 200) with,<br />
in white, their points of interest. Considering the density of points, their<br />
representation is difficult and the visual inspection of the images must be<br />
done carefully. Figure 2.4.2 shows matching points between these two<br />
images. An attentive observation shows that the matching is quite often<br />
correct. Obviously homologous points are in the overlap zones of the two<br />
images. Figure 2.4.3 presents two excerpts of these images, allowing one<br />
to see the details of the matching points in a small area of the images.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 2.4.2 Results of the matching.<br />
Automatization of aerotriangulation 129<br />
2.4.3.2 Calculations of homologous points<br />
• Photometric aspect: a similarity criterion much used in many matching<br />
techniques is the coefficient of correlation. This coefficient is calculated<br />
on two imagettes. The advantage of this technique resides in its<br />
ability to match two zones with similar grey levels. Nevertheless, it is<br />
necessary to use perfectly oriented images (problem of images with<br />
strong rotations).<br />
• Signal aspect: a criterion based only on the local resemblance (photometric<br />
criterion) is not always sufficient. Indeed, the correlation<br />
within a homogeneous zone or along a contour remains a source of<br />
ambiguity. A competing or complementary strategy consists in only
130 Franck Jung et al.<br />
Figure 2.4.3 Excerpts of the images with homologous points.<br />
matching the points possessing particular features of the signal (Harris<br />
and Stephens, 1988; Förstner and Gülch, 1987).<br />
A mixed approach takes full advantage of the two methods. The detection<br />
of points is made according to features of the signal and the matching<br />
is done according to a photometric criterion.<br />
2.4.3.3 Reliability<br />
Tools contributing to obtaining a good reliability of tie points are numerous:
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4111<br />
45111<br />
Automatization of aerotriangulation 131<br />
• Repeatability of the detector: one preferably has to use a point detector<br />
with a good repeatability. Examples are the detectors of Förstner<br />
(Förstner and Gülch, 1987) and Harris (Harris and Stephens, 1988),<br />
among which Schmid (1996) shows that Harris’s possesses the best<br />
properties of repeatability.<br />
• Multi-scale approach: this approach allows one to guide the research<br />
of the homologous point in full resolution with the help of a reliable<br />
prediction of the disparity on sub-sampled images. This approach<br />
allows two problems to be solved: on the one hand the problems of<br />
outliers error (one may limit the size of research zones voluntarily);<br />
on the other hand this approach allows a limitation of the combinatories<br />
of the problem of matching in full scale. One can note that the<br />
introduction of a digital terrain model (DTM) can partially replace<br />
the use of sub-sampled images to predict research zones at full scale.<br />
• Multiplicity: a visual assessment of the quality of tie points permits<br />
one to note that the percentage of erroneous tie points decreases<br />
appreciably with the order of the point (number of images for which<br />
the point is visible). There are nevertheless several ways to exploit this<br />
observation. A very reliable way is to consider a multiple point as<br />
valid if all associated measures are in total interconnection with regard<br />
to the similarity function, i.e. if each pair of measures is validated by<br />
a process of image by image matching. Figure 2.4.4 shows this mechanism:<br />
every couple of points of this multiple link has been detected<br />
as pairs of homologous points by the matching algorithm.<br />
It is necessary to note that the points with a strong multiplicity are necessary<br />
for the calculation of aerotriangulation.<br />
Photogrammetric filtering<br />
Forgetting any consideration of radiometric type, one can also use the<br />
photogrammetric properties of tie points. Indeed, each of these points are<br />
supposed to represent the same ground point. Therefore, all perspective<br />
rays must cross in one point. Several methods can be proposed to use this<br />
A<br />
Image 1<br />
B<br />
Image 2<br />
D<br />
Image 4<br />
C<br />
Image 3<br />
Figure 2.4.4 Example of multiple point of 4th order in total interconnection.
132 Franck Jung et al.<br />
property. The first consists in attempting to set up every couple individually,<br />
for example by the eight points algorithm (Hartley, 1997). Technically,<br />
this algorithm calculates the fundamental matrix associated to a couple of<br />
images. If one notes u and u′ the projective vectors representing, in the<br />
reference of the image, two homologous points, the fundamental matrix<br />
F is defined by: u′ T Fu 0. In return for some numeric precautions, the<br />
algorithm permits the calculation of the likeliest fundamental matrix associated<br />
to a sample of points. From the fundamental matrix, it is possible<br />
to determine the epipolar line associated to a point u, and therefore also<br />
to calculate the distance between u′, homologous of u, and this epipolar<br />
line. This distance is a sort of ‘residual’ thus making it possible to qualify<br />
points. This algorithm is very general, and permits a setting up of whatever<br />
is the configuration. It has been used therefore for all sets of two<br />
images having at least nine points in common, whatever their respective<br />
positions. In this practice, for every ‘couple’ one attempts an iterative stake<br />
in correspondence: one sets up, one calculates residues, one eliminates<br />
points having a residual greater than three rms, and one repeats the setting<br />
up, and this until this one is satisfactory, or the process is consolidated.<br />
If the convergence could not have taken place, one then tries a stochastic<br />
approach algorithm ([Zhang and Gang, 1996]): one chooses N samples of<br />
tie points and applies the algorithm on every sample. Residues are then<br />
calculated on the whole set of points, and one preserves the setting up<br />
where the median of the residues is the weakest. If one of the methods<br />
has given a satisfactory set-up, one eliminates all points whose residual is<br />
greater than three times the median.<br />
• If the number of exact points exceeds appreciably the number of errors,<br />
the setting up is valid, and the points whose discrepancy with the<br />
setting up is too high are eliminated. Another method is to really use<br />
the photogrammetric knowledge. While doing an approached setting<br />
up, errors appear. The only limitation is that most aerotriangulation<br />
software are not intended to manage thousands of points automatically<br />
generated by algorithms. Adapted tools must therefore be set up.<br />
• On terminating filtering, one wishes to generally choose among the<br />
remaining points, according to criteria like the distribution or the a<br />
priori validity. Indeed, the distribution of points is not absolutely<br />
uniform, some zones being over-equipped in comparison to others. In<br />
order to limit this phenomenon, and to limit the number of points in<br />
the aerotriangulation, it is necessary to choose among the set of filtered<br />
points. Criteria to take into account are the equidistribution of points,<br />
the a priori validity of these (in general, one can value the validity of<br />
points during their calculation), in order to guarantee having some<br />
exact points in every inter-images relation.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 2.4.5 (a) Examples of error sources (top: vehicle moving; bottom: areas<br />
locally similar).<br />
Figure 2.4.5 (b) Example of error source: textured zone.
134 Franck Jung et al.<br />
Figure 2.4.5 (c) Example of correct point.<br />
2.4.3.4 Precision<br />
To obtain precise tie points, two main approaches exist. The first consists<br />
in attempting to search for sub-pixel positions that maximize the function<br />
of similarity. This technique looks mostly like optimization methods. The<br />
important point concerning this method is that it links precision and resemblance:<br />
the tool used to put the points in correspondence and to localize<br />
them precisely is the same.<br />
The second is more geometric, and decouples questions of resemblance<br />
and precision. It aims at determining points corresponding to certain details<br />
of the ground in a precise way, without trying to put them simultaneously<br />
in correspondence (one will nevertheless note that the detection step is not<br />
completely independent from the setting in correspondence step; one should<br />
indeed preferably choose the points susceptible to being put easily into<br />
correspondence, namely those points with good geometric behaviour, like<br />
corners, crossings, etc.). A possible way consists in using the theoretical<br />
models of corners or positioning points on intersections of contours determined<br />
precisely. The putting into correspondence of these points will not<br />
change their localization.<br />
Generally speaking, the precision of localization of measure in images<br />
depends on the algorithm used for the detection of tie points. The impact<br />
of points of interest having a sub-pixel precision on the quality of the<br />
aerotriangulation is still, to a very large measure, to be proven. The<br />
matching aspect in this type of problem remains also very difficult. Subpixel<br />
coordinates of the homologous points can be estimated separately,<br />
or jointly during the phase of matching.<br />
For this approach, one can mention the techniques based on theoretical<br />
models of corners, or Blaszka and Deriche (1994), or Flandin (1999).
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 2.4.6 Example of multiple point of 4th order.<br />
2.4.3.5 Use of photogrammetry<br />
Even though the problem stands upstream of aerotriangulation, photogrammetry<br />
can constitute a considerable help in solving the problem of<br />
the measurement of tie points.<br />
<strong>Photogrammetry</strong> can guide methods (by a multi-scale approach) to<br />
reduce the combinatories of the problem (by the prediction of homologous<br />
positions at the time of the matching), and finally to filter the tie points<br />
after their calculation, which contributes appreciably to their reliability.<br />
2.4.4 Discussion<br />
Automatization of aerotriangulation 135<br />
The sections above have exposed some general considerations on objectives<br />
and methods bound to the automatic measurement of tie points. We<br />
have also to discuss limits and extensions of the existing processes. Indeed,<br />
even though the existing systems give good results, there are many problematic<br />
cases, as well as numerous possible extensions to less classic cases.<br />
These problems may be analysed according to several axes intervening in<br />
photogrammetric cases: the sensor, the scene, the image acquisition.<br />
Finally, we will present some algorithmic considerations.
136 Franck Jung et al.<br />
2.4.4.1 The sensor<br />
Use of colour<br />
A priori, the use of colour should allow for a better identification of ground<br />
details, but currently it proves to be that most methods use mainly images<br />
in grey levels, for the detection of points of interest and for the calculation<br />
of tie points as well.<br />
Quality of the signal<br />
One expects a priori to get tie points of better quality with a sensor of<br />
better properties: detectors of points of interest generally use the differential<br />
properties of the signal; they are therefore sensitive to the image<br />
noise. Thus, a better signal to noise ratio in images draws a better stability<br />
from these operators. In practice, the evaluation of the sensitivity to the<br />
sensor requires the use of important equipment, so as to perform flights<br />
in identical conditions, so that no survey has been conducted on the question<br />
up to now.<br />
Multi-sensor data<br />
The integration of multi-sensor data is a problem that will occur, because<br />
one can consider embarking several cameras simultaneously for a flight<br />
(for example, simultaneous embarking of sensors specialized by channel:<br />
red, green, blue, and even infrared). No survey on this question and no<br />
software dealing with this problem in an operational condition is currently<br />
available.<br />
2.4.4.2 The scene<br />
Variation of the aspect of objects<br />
The aspect of objects can vary according to the point of view. In the case<br />
of a slope, the perspective distortions can be appreciable (the worse case<br />
is facades in vertical aim). The occlusions are not the same. Finally, the<br />
aspect can change according to the point of view. Figure 2.4.7 presents<br />
two images of the same scene, obtained during the same flight, from two<br />
different viewpoints. The superposition of the two images shows distinctly<br />
the variations of intensity observed from these two points of view.<br />
Forests, shades, textures<br />
These objects generally create problems in image analysis. In the case of<br />
the measure of tie points, they provoke three types of problems. Trees
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Automatization of aerotriangulation 137<br />
Figure 2.4.7 Example of change of aspect according to the point of view.<br />
cause problems of precision of pointing. Shades move between two image<br />
acquisitions, notably between two strips (the interval of time between two<br />
neighbouring images of two different strips can exceed 1 h), and therefore,<br />
even though extremities of shades can seem easy to delineate, they<br />
are not valid from a photogrammetric point of view (see Figure 2.4.8).<br />
Textures naturally lead to problems of identification: it is easy to confuse<br />
two very alike details in zones with repetitive motives, typically: passages<br />
for pedestrians.<br />
Relief<br />
The relief provokes notable changes of scale and variations of disparity.<br />
When looking for homologous points, the space of research is generally<br />
larger. In this case, the utilization of photogrammetry on the under-sampled<br />
images (multi-scale approach) can prove to be useful.<br />
Case of terrestrial photogrammetry<br />
In this case, often the hypothesis of a plane land (and horizontal) is no<br />
longer valid. The space of disparities is more complex than in the aerial<br />
case (in the aerial case, with little relief, two images of the same scene are<br />
indeed nearly superimposable, which is not the case in terrestrial imagery).<br />
Figure 2.4.9 presents the case of a terrestrial scene where a street axis is<br />
parallel to the optic axis: distortions bound to the perspective are very<br />
significant, as well as the disparities of points.
138 Franck Jung et al.<br />
Figure 2.4.8 Example of shade point extremity.<br />
Figure 2.4.9 Terrestrial image acquisition example: the perspective distortions can<br />
be significant.<br />
2.4.4.3 The image acquisition<br />
Rotation of images<br />
The interest in using images of any rotation is to achieve some diagonal<br />
strips in photogrammetric blocks, which makes them more rigid. Otherwise,<br />
if one tries to calculate two overlapping photogrammetric studies<br />
simultaneously, the case of large rotations may occur. Techniques of matching<br />
generally use linear correlation coefficients between stationary windows.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
In the case of known rotation between images, these can be sampled so<br />
that these techniques work. In the case of unknown angle rotation, it is<br />
necessary to adapt the methods. Invariants by rotation resemblance<br />
measures have been developed: one can consider turning the imagettes<br />
(Hsieh et al., 1997), or to consider measures relying directly on invariants<br />
(Schmid, 1996).<br />
Image acquisitions at various scales<br />
In the case of image acquisitions of different scales, measures of resemblance<br />
must also be adapted, because they don’t generally support strong<br />
changes of scale. One however knows the scales of image acquisitions,<br />
which reduces the problem. But even at known relative scales, the problem<br />
is significant, because in order to compare two imagettes of two different<br />
scales, it is necessary to adapt one of the signals. Otherwise, for example<br />
for the positioning of points of interest, errors committed by the detectors<br />
may be different between two scales.<br />
Image acquisitions at different dates<br />
If one wishes to process simultaneously some images taken at different<br />
dates, it is necessary to be careful after the inversions of contrast that may<br />
occur (notably by reason of different positions of the shades). In the same<br />
way, the aspect of the vegetation can change. Considering the evolution<br />
of the landscape, a certain number of topographic objects can undergo an<br />
important radiometric evolution (ageing of roofs).<br />
2.4.4.4 Algorithmic considerations<br />
Automatization of aerotriangulation 139<br />
Parametrage of methods<br />
As in all processes of image analysis, the survey of the parameters must<br />
be made in order to learn how to master the method and to know its<br />
limits. One will for example aim at reducing the number of critical parameters:<br />
one can easily replace a threshold level of a correlation coefficient<br />
(that is generally very sensitive) by the introduction of the symmetrical<br />
crossed correlation (see description of homologous points in §2.4.1.2).<br />
Number of useful points by image<br />
Studies have shown that there was no improvement to the result of aerotriangulation<br />
by using more than 20 points by image (if the 20 points are<br />
correct). In the case of automatic measure, it is necessary to have more<br />
points to be able to reject reliably the aberrant points. The necessary point<br />
number is therefore higher (see §2.4.2.3).
140 Franck Jung et al.<br />
Evaluation<br />
The assessment of the quality of tie points is based on the quality of the<br />
aerotriangulation resulting from this setting up. A first criterion of quality<br />
is the values of the residues (in microns or in pixels) of tie points. Another<br />
technique consists in assigning values to the quality of tie points with the<br />
help of reference ground points. These ground points can be calculated<br />
with the help of a reliable aerotriangulation or with ground measures.<br />
Combinative<br />
At the time of the research of homologous points between two images,<br />
one may a priori attempt to put in correspondence any point of the first<br />
image, with any point of the second image. In this case, the combinative<br />
of the problem is important. In the case of n images, the combinative<br />
increases again. The combinative can be reduced strongly while restricting<br />
correctly the space of research of points homologous of a given point.<br />
A classical technique for it is the multi-scale approach: by doing first<br />
some calculations at a reduced scale, one may predict the position of a<br />
point to search for.<br />
Another solution resides in the ‘geometric’ approach: if, for a given<br />
research zone, one does consider as good candidates only the positions of<br />
points of interest that are present, and not all pixels of the region, then<br />
the combinative falls appreciably. However, this method rests entirely on<br />
the repeatability of the detector of points of interest.<br />
2.4.5 Automatic constitution of an index map<br />
The objective of the constitution of an index map is to calculate 2D information<br />
of position and orientation of images of the flight, permitting the<br />
placing of all these images in the same reference mark of the plan, so that<br />
the homologous details of several images are at the same position in this<br />
common reference frame. For two images possessing an overlap, it is then<br />
possible to appraise a change of reference frame permitting the forecasting<br />
of the position of a homologous point to a given point p. Thus, the index<br />
map, beyond its own utility, serves as a guide for the detection of multiple<br />
tie points.<br />
Images are treated after a strong sub-sampling. One considers a global<br />
reference frame to the plane, common to all images. It is the reference of<br />
the block. What one searches for is, for every image I, a similitude permitting<br />
the passage of coordinates of a point of the plane (x B , y B ) expressed<br />
in the reference of the block, to its position p(x, y) in the image. The<br />
problem is treated strip by strip, then by an assembly of strips. Techniques<br />
rely on the detection of points of interest in images, and on the calculation<br />
of the relative 2D orientation of a couple of images whose relative<br />
rotation is an angle close to 0.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Automatization of aerotriangulation 141<br />
Detection of points of interest<br />
This part of the method includes two parameters: the standard deviation<br />
of the Gaussian used for smoothing and derivations, and the size of the<br />
calculation window associated to these processes.<br />
Processing a couple of neighbouring images<br />
One supposes that neighbouring images possess a relative rotation of angle<br />
of nearly 0, and that because of the strong sub-sampling, the disparities<br />
of homologous points of the two images are all similar. Two homologous<br />
points can be put therefore in correspondence without the problem of rotation.<br />
For two points of interest Pi of the image I, and Qj of the image J<br />
one considers cij the coefficient of calculated linear correlation on a square<br />
window centred on each point. Two points Pi and Qj are judged homologous<br />
candidates if for Pi no point of J is more alike, and reciprocally.<br />
Formally:<br />
∀j′, c ij′ c ij, ∀i′, c i′j c ij . (2.15)<br />
One thus gets couples of points candidates for the images I and J. In practice,<br />
many aberrations are present. One operates therefore a filtering in<br />
the space of disparities. For two homologous candidate points, one defines<br />
the vector of disparity joining these two points. This is the vector obtained<br />
by difference of coordinates of points. While accumulating these disparity<br />
vectors in a histogram, one gets a cloud corresponding to the points of<br />
similar disparities. These points constitute the homologous points retained<br />
by the method.<br />
In practice, the cloud is identified by a morphological closing with an<br />
square structuring 3 × 3 element. After closing, the connex component of<br />
the strongest weight is identified, then the points participating in this<br />
component are selected. At this step one defines the performance of an<br />
image couple: it is the ratio between the number of points participating<br />
in the cloud of points and the number of homologous candidates points.<br />
From these points one appraises with the least squares the likeness<br />
allowing the best joining of the two references of the images. It permits<br />
the second image to be put in the reference frame of the first. This part<br />
of the method includes only one parameter concerning the correlation: it<br />
is the size of the calculation window. The size of the structuring element<br />
for filtering is otherwise stationary.<br />
Process of a strip<br />
One supposes one knows N images constituting a strip. The previous<br />
process is applied to the N1 successive couples constituting a strip. One<br />
can thus ‘sink’ all images of a strip in the reference of the first image of
142 Franck Jung et al.<br />
this strip. All strips are treated independently. There is no supplementary<br />
parameter concerning this step.<br />
Process of a block<br />
One supposes one knows the order of strips. One also supposes that strips<br />
possess a correct lateral overlap (> 10 per cent). This part of the method<br />
is achieved by iterations: one glues the strip n 1 to the organized block<br />
of the n previous strips. One considers therefore that one already has a<br />
block organized to which one adds a supplementary strip. It is achieved<br />
while first hanging an image of the strip to the block, then hanging up<br />
the other images of the strip, image by image.<br />
• To hang up an image of the strip to the block: for an image I of the<br />
strip, one temporarily considers all couples between the images I and<br />
J of the last strip of the block. One also considers couples between<br />
the images I and J having undergone a rotation of . For all formed<br />
couples, one performs the process of image couples studied earlier,<br />
then one keeps the couple possessing the best performance. If this<br />
performance is not sufficient, one repeats the operation until one finds<br />
an effective enough couple. One thus gets a couple in an image I 0 and<br />
J 0 permitting the image I 0 to be put in the reference frame of the block.<br />
• Progressively one sinks images of the strip in the reference of the<br />
couple. To avoid error accumulations, an optimization of the system<br />
is done at regular intervals (every n opt images), by a method described<br />
in the following.<br />
This part of the method introduces two parameters. The first is the<br />
threshold of performance beyond which one accepts the coupling of image<br />
between the strip and the block. This threshold is fixed to a value that a<br />
precise survey showed was quite reliable. The second is n opt 5. This<br />
value allows the method to be reliable while reducing the number of optimizations.<br />
Optimization of block<br />
This optimization aims at refining the positions and orientations of the<br />
images of a block. This is necessary because it is not possible of ‘to glue’<br />
in a rigid way two strips because of distortions that accumulate at the<br />
time of the constitution of the intra-strip index map. One considers all<br />
couples of images possessing an overlap. For each couple, the homologous<br />
points to the previous sense are calculated. A function of cost aiming to<br />
achieve the objective of the image assembly is minimized. Technically this<br />
function is the sum of cost functions bound to every couple of images.<br />
The function bound to a couple of images aims to minimize, for each of
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Automatization of aerotriangulation 143<br />
the homologous points of the image couple, the distance of the two homologous<br />
in the reference of the assemblage image. This function is adjusted<br />
so as to avoid the aberrant points.<br />
Survey of parameters for the constitution of the index map<br />
The behaviour of parameters is known. It appears notably that for some<br />
among them, values are stationary in the sense where meaningful studies<br />
have been processed without modification. These are:<br />
• the value of the standard deviation of the Gaussian in the calculation<br />
of points of interest, as well as the size of the associated calculation<br />
window;<br />
• the size of the correlation window for the calculation of homologous<br />
candidates points;<br />
• the size of the structuring element used in filtering in the space of<br />
disparities.<br />
For the other parameters, their behaviour is known: their variation allows<br />
an increase in reliability at the cost of more calculation time in any case,<br />
or inversely. In this sense, the sensibility of the parameters is low as their<br />
influence on the results is identified.<br />
Figure 2.4.10 Excerpt of an index map (city of Rennes, digital camera of IGN-F).<br />
14 strips of 60 to 65 images. The geometric precision of<br />
mosaicking is good: the errors on the links range to 1 or 2 pixels.
144 Franck Jung et al.<br />
2.4.6 Conclusion<br />
The problem remains wide open. Nevertheless, under certain conditions,<br />
some systems are quite capable of producing tie points adapted to calculate<br />
photogrammetric works. These tools can be very useful for large<br />
studies.<br />
In any case, digital photogrammetry will need such tools, which justifies<br />
completely the ongoing developments in this domain: the problematic<br />
zones (occlusions, strong relief, texture, change of date, strong scale variation,<br />
etc.) remain difficult points for these methods.<br />
References<br />
Blaszka T., Deriche R. (1994) Recovering and characterizing image features using<br />
an efficient model-based approach. Research Report from INRIA, Roquencourt,<br />
France, no. 2422.<br />
Flandin G. (1999) Détection de points d’intérêt sub-pixellaires pour l’orientation<br />
relative 3D de images. Stage Report, Institut Géographique National France.<br />
Förstner W., Gülch E. (1987) A fast operator for detection and precise location<br />
of distinct point, corners and centre of circular features. Proceedings of the<br />
Intercommission of the International Society for <strong>Photogrammetry</strong> and Remote<br />
Sensing, Interlaken, Switzerland, pp. 281–305.<br />
Harris C., Stephens M. (1988) A combined edge and corner detector. Proceedings<br />
of the 4th Alvey Conference, Manchester, pp. 147–151.<br />
Hartley R.I. (1997) In defense of the Eight-Point Algorithm. IEEE Transactions<br />
on pattern analysis and machine intelligence, vol. 19, no. 6, June, pp. 580–593.<br />
Heipke C. (1998) Performance of tie-point extraction in automatic aerial triangulation.<br />
OEEPE Official publication, no. 35, pp. 125–185.<br />
Heipke C. (1999) Automatic aerial triangulation: results of the OEEPE-ISPRS test<br />
and current developments. Proceedings of Photogrammetric Week, Institute for<br />
<strong>Photogrammetry</strong>, Stuttgart, pp. 177–191.<br />
Hsieh J.W., Liao H.Y.M., Fan K.C., Ko M.T., Hung Y.P. (1997) Image registration<br />
using a new edge based approach. Computer Vision and Image Understanding,<br />
vol. 67, pp. 112–130.<br />
Schmid C. (1996) Appariement d’images par invariants locaux de niveaux de gris.<br />
Thèse de doctorat de l’institut National Polytechnique de Grenoble.<br />
Zhengyou Zhang, Gang Xu (1996) Epipolar Geometry in Stereo, Motion and<br />
Object Recognition. Kluwer Academic Publishers, pp. 102–105.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
<strong>Digital</strong> photogrammetric workstations 145<br />
2.5 DIGITAL PHOTOGRAMMETRIC WORKSTATIONS<br />
Raphaële Heno, Yves Egels<br />
2.5.1 Introduction: the functions of a photogrammetric<br />
restitution device<br />
Whatever its principle of realization, one will find in any photogrammetric<br />
restitutor a certain number of basic functions that one can represent on a<br />
working diagram (Figure 2.5.1). This principle of working was discovered<br />
during the first industrial realizations at the beginning of the twentieth<br />
century (restitutor of Von Orel at Zeiss, Autograph of Wild). In the<br />
first systems, the geometric function is achieved by an optic or mechanical<br />
analogue computer (whence their ‘analogical restitutors’ name). The<br />
two other functions use a human operator, whose stereoscopic vision<br />
assures the image matching, and whose cultural knowledge the interpretation.<br />
In the 1970s, a first appreciable evolution appeared, the analytic restitutor:<br />
image formulas are calculated analytically by a computer, which<br />
displaces images from an optic system using a servo, because images being<br />
Movement<br />
Input<br />
peripheral<br />
No<br />
No<br />
Left image<br />
formula<br />
Field<br />
x, y, z<br />
Right image<br />
formula<br />
Left x, y Right x, y<br />
Are the<br />
points<br />
homologous?<br />
Yes<br />
Is<br />
the point<br />
interesting?<br />
Record of the point (GIS)<br />
Figure 2.5.1 Schematic diagram of the photogrammetric restitution.<br />
Yes<br />
Geometry<br />
Correlation<br />
Interpretation
146 Raphaële Heno and Yves Egels<br />
on films, the matching as well as the interpretation are always done by a<br />
human operator.<br />
Today, images become digital, and can be visualized on a simple<br />
computer screen. Besides, non-geometric functions are more often taken<br />
into account by algorithmic means, and the human operator, and therefore<br />
the visualization itself, then becomes superfluous. In this last case,<br />
there isn’t a photogrammetric restitutor strictly speaking, but only a specialized<br />
software functioning on more or less powerful computers.<br />
In the following, one is concerned mainly with the case where a human<br />
operator must, instead of achieving the totality of photogrammetric process<br />
by hand, at least supervise it, control its results, and possibly perform<br />
certain operations by hand.<br />
2.5.2 The technology<br />
2.5.2.1 The equipment<br />
The passage to the digital in photogrammetry allowed a freedom from the<br />
high-precision optic and mechanical components, which kept the acquisition<br />
and maintenance costs of systems at a very high level. In addition,<br />
the size of the machines was considerably reduced (see Figure 2.5.2).<br />
Systems of digital photogrammetry are based on workstations or, increasingly,<br />
on PCs. The recent computers include almost the whole of the<br />
necessary means for photogrammetric restitution, and today have enough<br />
power for these applications. Nevertheless it is necessary not to forget<br />
that the manipulated data images generally have sizes of several hundreds<br />
of megabytes, and that a given study will require the use of tens, or even<br />
hundreds of images. The capacity of storage, and access times, should<br />
therefore be especially adapted.<br />
Only two indispensable elements are not yet completely standard, even<br />
if the development of video games makes them more and more frequent:<br />
the peripheral for data input and the stereoscopic visualization; as regards<br />
the computer conception, digital photogrammetry is hardly more than a<br />
video game with a technical aim.<br />
Peripheral for data input<br />
The system must possess a peripheral of coordinate input able to address<br />
three independent axes. Solutions are numerous, from arrows on the keyboard,<br />
to the cranks and pedal of traditional photogrammetric devices, or<br />
controllers of immersion of virtual-reality stations, joysticks, or the specialized<br />
mouse. Only the cost/ergonomics ratio can overrule them. The practice<br />
establishes that the movement must be as continuous as possible, and<br />
that the control in position is far preferable to speed commands (joysticks).<br />
When a specialized peripheral is developed, supplementary buttons may be
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
<strong>Digital</strong> photogrammetric workstations 147<br />
usefully attached, allowing one to simplify the use of the most frequent<br />
commands.<br />
Stereoscopic visualization<br />
The system must also allow stereoscopic vision, inasmuch as one will ask<br />
the operator to achieve some visual image matching. Here also, several<br />
solutions exist, and some are under normalization. It is about presenting<br />
to each eye the image he has to observe. In order of increasing comfort,<br />
one can mention:<br />
• the display of the two images in two separated windows, which the<br />
operator will observe with an adequate optic system of stereoscopic<br />
type, placed before the screen;<br />
• the simultaneous display of the two images, one in green, the other<br />
in red, the operator benefiting from the structural complementarity of<br />
such glasses (anaglyphs) whose cost is extremely low;<br />
• the alternative display of the two images, a device using liquid crystals<br />
letting the image pass toward the eye to which it is destined.<br />
In the professional systems, only this last solution is used, due to its better<br />
ergonomics. The frequency of the screen conditions the quality of the<br />
stereoscopic visualization, because the alternative display divides by two<br />
the frequency really discerned by the user (a minimal frequency of 120<br />
Hz is necessary to avoid operator fatigue). Several convenient realizations<br />
are susceptible of being used. The liquid crystals may be placed directly<br />
before the eye, a low-cost solution, synchronized by a wire or by infrared<br />
link. Or they may be placed on the screen, the operator then using a couple<br />
of passive polarized glasses (which allows him or her to look indifferently<br />
Control screen<br />
Cranks and<br />
pedal<br />
commands<br />
Stereo display<br />
Figure 2.5.2 Diagram of a digital photogrammetry system.<br />
Hard disks<br />
Central unit
148 Raphaële Heno and Yves Egels<br />
at the stereoscopic screen and at the command screen), the most comfortable<br />
solution but also the most expensive.<br />
2.5.2.2 Photogrammetric algorithms<br />
It was not necessary to reinvent photogrammetry to make it digital. The<br />
equations are the same as those on the analytic photogrammetry systems<br />
(equations of collinearity or coplanarity).<br />
More and more systems are ‘multi-sensor’, that is to say that they can<br />
not only process aerial images, but also different geometry images, for<br />
example the images from scanning sensors or from radar. The setting up<br />
of these images uses mathematical models different from the traditional<br />
photographic perspective, possibly parametrable by the user.<br />
Whatever their origin, once images are set up or georeferenced, the function<br />
‘terrain → image’ allows one to transmit in real time the displacement<br />
terrain introduced with the mouse or the cranks.<br />
On the other hand, contrary to the analogical or analytic restitutors,<br />
the availability of the image under digital shape allows a considerable<br />
spread of the possibilities of the automation of photogrammetry, requiring<br />
to bring together in one unique system photogrammetric functions and<br />
functions of image processing and shape recognition.<br />
2.5.3 The display of image and vector data<br />
2.5.3.1 Display of images<br />
The used images always have some sizes much greater than the dimension<br />
of the screen (often 100 times larger). It will thus be necessary to be able<br />
to displace the display window conveniently within the total image. Two<br />
modes of displacement are possible: either the image is fixed, and the cursor<br />
of measure mobile; or the cursor is fixed and central, and the image is<br />
mobile. The measure will be performed by superimposing a pointing index<br />
(whose colour will be adjusted automatically to the background image).<br />
The image is fixed<br />
The fixed image/mobile cursor configuration is easiest to implement: it is<br />
just a question of displacing some octets in the video memory. But the<br />
ergonomics of this solution is mediocre. If images are not reprocessed<br />
geometrically (epipolar resampling), the transverse parallax is eliminated<br />
in only one point, generally the centre of the image. During the displacements<br />
of the cursor, some parallax appears, which makes the pointing if<br />
not impossible, at least imprecise. Besides, when the cursor reaches the<br />
side of the screen, the system must reload another zone of images, which<br />
interrupts the work of restitution and distracts the operator.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
<strong>Digital</strong> photogrammetric workstations 149<br />
The cursor is fixed<br />
More demanding in calculation capacities, the mobile image/central cursor<br />
configuration is far more preferable. If the displacement of images is fluid<br />
enough, and it must be so not to tire the operator, this configuration recalls<br />
the analytic devices, where image-holders move in a continuous way in<br />
front of a fixed mark. Images are generally charged in memory by tiles,<br />
to optimize the time of reloading.<br />
The system sometimes proposes a global view (under-sampled image)<br />
for the fast displacements in the model.<br />
In the ‘exploitation of the model’ mode, it is preferable to dedicate the<br />
maximum surface of the stereo screen to the display of the two images.<br />
But then, at the time of aerotriangulation points measurement, it is convenient<br />
to be able to display for every point all images where it is present<br />
(multi-window).<br />
2.5.3.2 Zooms and subpixel displacement<br />
Zooms by bilinear interpolation, or to the nearest neighbour, serve to<br />
enlarge all or part of the current screen. Their goal is usually to allow to<br />
point with a better precision than the real pixel of the image. Unfortunately,<br />
the zoom decreases the field of observation simultaneously and adds a<br />
significant fuzziness to the image, these two effects degrading appreciably<br />
the quality of the interpretation. Subpixel displacement of the image allows<br />
these two defects to be palliated simultaneously (but this function is not<br />
generally available in the standard graphic libraries, and requires a specific<br />
programming).<br />
In the case where the image is fixed, it is enough to resample the cursor<br />
so that it appears positioned between two pixels, which requires that it is<br />
1/1<br />
Figure 2.5.3 Rear zoom pyramid.<br />
1/2<br />
1/4<br />
1/16<br />
1/8<br />
1/32
150 Raphaële Heno and Yves Egels<br />
even-formed of several pixels. If the image is mobile, then it is the totality<br />
of the displayed image that it is necessary to recompute.<br />
Concerning the rear zooms, they are immediately calculated or generated<br />
in advance under the shape of a pyramid of physical images of<br />
decreasing resolution (see Figure 2.5.3). In the case of image pyramids,<br />
one won’t be able to zoom back at any scale.<br />
2.5.3.3 Image processing<br />
Within basic tools one finds at least the functionalities of adjustment of<br />
the contrast, the brightness, of positive/negative inversion. Based on the<br />
colour table (LUT, a term that stands for look-up table) of images, they<br />
work in real time, and don’t require the creation of new images.<br />
It is sometimes possible to apply convolution filters (contours improvement,<br />
smoothing . . .). Their application being quite demanding for the<br />
CPU, it slows down considerably the time required for image loading.<br />
Besides, these filters often have some perverse effects on the geometric plan<br />
(displacement of the contours), and the influence on the precision of the<br />
survey can be catastrophic. For a better efficiency, one will prefer sometimes<br />
to calculate new images.<br />
2.5.3.4 Epipolar resampling<br />
The visualization of a stereoscopic model whose images have been acquired<br />
with appreciably different angles can be tedious for operators, because the<br />
two images have too different scales and orientations. Performances of<br />
usual image-matching techniques are degraded also by this type of images.<br />
(See Figure 2.5.4.)<br />
In this case, it is desirable to make an epipolar resampling: one calculates<br />
the homographic plane that allows one to pass from the initial image<br />
Image acquisition Left image Right image<br />
Object Homologous<br />
object<br />
Figure 2.5.4 Visualization of stereoscopic model using images with significant<br />
differential rotations.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Images resampled in epipolar geometry<br />
<strong>Digital</strong> photogrammetric workstations 151<br />
Object Homologous<br />
object<br />
Epipolar lines<br />
Figure 2.5.5 Epipolar resampling.<br />
Plane containing<br />
the basis<br />
to the one that would have been obtained if optical axes of photos had<br />
been parallel between them and perpendicular to the basis (normal case);<br />
after this process, two points of a line of the left image will have their<br />
counterparts in one line of the right image, and the transverse parallax is<br />
constant. Epipolar lines are the intersections of the bundle of planes<br />
containing the basis and of the two image planes. In the resampled images,<br />
these lines are parallel, whereas they were converging in the initial images.<br />
(See Figure 2.5.5.)<br />
This resampling requires at least the calculation of the relative orientation<br />
(and of course cannot be used at the same time as the measures<br />
necessary for setting it up). It generates two new images, which must be
152 Raphaële Heno and Yves Egels<br />
taken into account in the management of the disk space. As it is valid for<br />
only one couple, it is possible to cut up images and to keep only the<br />
common parts of images of the model.<br />
Epipolar resampling not only improves the operator’s comfort, but it<br />
accelerates the work of correlators, since the zone of research of the homologous<br />
points is reduced to a one-dimensional space (instead of a twodimensional<br />
space for a systematic correlation of images). Some systems of<br />
digital photogrammetry require working with images in epipolar geometry<br />
for the extraction of the DTM; others do this resampling continuously.<br />
2.5.3.5 Display of the vectors<br />
Whereas on the analytic restitutors the addition of a vector display on<br />
the images (very useful for cartographic updating) was a very costly<br />
option, requiring complex optical adaptations, the superposition of vectors<br />
in colour to the images in stereo doesn’t pose any problem for digital<br />
photogrammetry systems.<br />
2.5.4 Functionalities<br />
<strong>Digital</strong> restitution systems offer at least the same functionalities as the analytic<br />
restitutors. Besides, these can practically be reused without change,<br />
except for the addition of functions to help pointing, which profit from the<br />
digital nature of the image. But their features allow one to consider numerous<br />
extensions of their use. We summarize here the main among them.<br />
Contrary to what happened in analogical or analytic systems, the<br />
geometric quality of the products calculated on the digital photogrammetry<br />
system (vector data base, DTM, orthophotos) will be reasonably<br />
independent of the system. The geometric algorithms used are very close,<br />
and no mechanics intervenes. The only difference lies in the pointing capability<br />
which may be at the pixel level, or using sub-pixel methods. One<br />
should especially be concerned with the quality of the following steps:<br />
• condition of the films;<br />
• quality of the digitization (type of scanner used, control of the parameters,<br />
scanner resolution);<br />
• quality of the control points;<br />
• quality of measures of points of relative and absolute orientation.<br />
2.5.4.1 The management of data<br />
The manipulated images<br />
<strong>Digital</strong> photogrammetry workstations are capable of working with images<br />
originating from scanned classic aerial pictures, scanned terrestrial images,
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
<strong>Digital</strong> photogrammetric workstations 153<br />
or with the images of aerial or spatial digital cameras (e.g. the CCD matrix<br />
digital camera of the IGN-F, the DMC of Zeiss Intergraph, the APS 40<br />
of Leica Helava Systems, see §1.5).<br />
The black and white 8 bits and colour 24 bits images are easily read<br />
by digital photogrammetry systems. Considering that the images may have<br />
12 bits or more by channel that are meaningful for the digital sensors, it<br />
happens that only the first 8 bits of every pixel are used. Anyway, the<br />
current computer screens actually display 6 bits only, which is also the<br />
true dynamics of movies, and the human eye only distinguishes a part of<br />
the 256 levels of grey of a coded image on 8 bits, but with the large<br />
dynamics digital sensors, this loss of information is prejudicial for algorithms<br />
of image matching and automatic shape recognition.<br />
There is no consensus within manufacturers of digital photogrammetry<br />
systems on the computer format of images to use. Nearly all are capable<br />
of reading and writing the TIF (tiled or not), but some recommend<br />
converting it to a proprietary format to optimize process times. It is the<br />
same for the JPEG compression: its direct utilization slows down some<br />
applications, as this technique (in its present state) prevents the direct<br />
access to portions of image. The wavelet compression gives excellent results<br />
in terms of the ratio ‘quality of compressed image/reduction of volume’,<br />
and it is likely that the digital photogrammetry systems will quickly adopt<br />
this process.<br />
Photogrammetric database<br />
The data necessary to important studies exploitation are quite numerous<br />
and voluminous; they require a rigorous organization that is taken into<br />
account, either by a hierarchy predefined by directories, or by a specific<br />
database.<br />
Systems generally allow one to store data by project, that correspond<br />
intuitively to a given geographical zone, to a theme, etc. A project can<br />
correspond to a directory of the arborescence of the system, in which one<br />
finds the files necessary for setting up the models (files of camera, files<br />
containing reference points, possibly files containing the measures on the<br />
image achieved on another device . . .), as well as files generated by calculations<br />
(parameters of the internal orientation, position and attitudes of<br />
the camera for each of the cliches, matrixes of rotation . . .). Data are<br />
managed by model, or image by image (a file by model, or a file by image).<br />
Images are stored in this same directory, or, for convenience in the management<br />
of the disk’s space, merely referenced there. (See Figure 2.5.6.)<br />
The local networks (Ethernet 100, then Giga Ethernet) allow work in<br />
architecture on a customer-server basis: images are set up and stored on<br />
a machine ‘server’, while machines’ ‘customers’ use them (visualization,<br />
process, survey) via the network, which requires mechanisms of control<br />
of the data consistency.
154 Raphaële Heno and Yves Egels<br />
Data in input<br />
Data in output<br />
images<br />
2.5.4.2 Basic photogrammetric functionalities<br />
Setting up of a stereoscopic couple<br />
INTERNAL ORIENTATION<br />
project<br />
101.img<br />
103.img<br />
105.img<br />
107.img<br />
Camera.fil<br />
Reference.fil<br />
101 – 103.fil<br />
103 – 105.fil<br />
105 – 107.fil<br />
Possible content<br />
Figure 2.5.6 Example of photogrammetric database.<br />
• Directory of model images<br />
• Operator parameters<br />
• Focus<br />
• Parameters of the internal orientation for each<br />
image<br />
• Position of the camera for each image<br />
ω, φ, κ, X, Y, Z<br />
The internal orientation phase is always necessary for images coming from<br />
a digitized classic aerial photograph. Knowledge of the theoretical position<br />
of reference marks on the bottom plate of the camera (via the certificate<br />
of calibration of the camera) and the shape of these reference marks allowed<br />
at least a partial automation of this phase of measures, thanks to image<br />
matching.<br />
The operator is sometimes asked to search by hand for the first reference<br />
marks, the system recognizing the next ones automatically. Automatic<br />
image matching, if necessary, assists interactive pointing.<br />
The internal orientation is completely automatic on certain systems.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
EXTERNAL ORIENTATION<br />
Most systems calculate the relative and absolute orientations by the simultaneous<br />
use of the collinearity equation. Even though the measure of points<br />
of relative orientation is assisted by image matching, or even completely<br />
automatic in cases where one uses a theoretical distribution of points, it<br />
is always imperative that an operator marks by hand the position of reference<br />
points.<br />
The elimination of the parallax in X and in Y is generally performed<br />
by blocking one of the two images, and by moving the other with the<br />
positioning system (cranks, mouse). Once the choice of the fixed image is<br />
made, one can ask for a pointing by image matching. In certain cases, this<br />
one can fail (vegetation, periodic textures, homogeneous zones); the use<br />
of the image matching does not dispense with a pointing by the operator,<br />
even if it is not so precise, because it remains more reliable in spite of<br />
everything.<br />
As soon as enough points are measured (six points to have residues of<br />
relative orientation, three ground points), one can launch a first calculation.<br />
Then the function ‘ground → image’ is known, and allows one to<br />
move automatically, for example on a supplementary control point. When<br />
measures are sufficiently overabundant, the examination of calculation<br />
residues allows one to control the quality of the setting up, and to validate<br />
it.<br />
Measures and acquisition in a stereoscopic couple<br />
The acquisition is very often limited to a stereoscopic couple; the automatic<br />
loading of images of the neighbouring couple on the edge of the<br />
image is sometimes possible, but of course requires having all images<br />
concerned already loaded on the device.<br />
Basic functionalities of the digital photogrammetric system are:<br />
• acquisition of points, lines, surfaces;<br />
• destruction, edition of these structures.<br />
<strong>Digital</strong> photogrammetric workstations 155<br />
Ideally, the digital photogrammetric system is interfaced with a GIS that<br />
manages data acquisitions in a database.<br />
All functionalities of the GIS (spatial analyses, requests, multi-scale<br />
cartography) are then directly usable after the data acquisition. Data are<br />
structured according to a more elaborate topology (line, spaghetti, sharing<br />
of geometry, object structure).<br />
Data extracted from the digital restitutors are systematically tridimensional.<br />
But rare are yet the GIS that allow one to manage completely a ‘true<br />
3D’ topology (management of vertical faces, of overhangs, of clearings),<br />
and not only of the 2.5D (planar topology, only one Z attribute by point).
156 Raphaële Heno and Yves Egels<br />
The GIS generally propose interfaces toward the most current formats<br />
(DXF, DGN, Edigeo . . .).<br />
2.5.4.3 Advanced photogrammetric functionalities<br />
Aerotriangulation<br />
Aerotriangulation allows one to georeference simultaneously all images of<br />
a site, using as much as possible the overlaps that they have in common,<br />
with a minimum number of reference points. This operation starts with a<br />
phase of measure of the image coordinates of a certain number of points<br />
seen on the largest number possible of images. Then a calculation in block<br />
allows one to determine the set of photogrammetric parameters of the<br />
work. The modules of available aerotriangulation calculation on systems<br />
of digital photogrammetry use the same formulas as the analytic aerotriangulation,<br />
and are most of the time identical.<br />
The measure of reference points, always interactive, is helped by the<br />
multi-windowing: once a point is measured on a image, the system can<br />
display in mini-windows all images susceptible of containing it, directly<br />
zoomed on the zone concerned. This is made possible by the a priori<br />
knowledge of the overlaps between images and their position in relation<br />
to each other, determined by GPS, or via the assembly table defined by<br />
the operator. The only work remaining for him is to measure the point<br />
position in the window in which it is present, in monoscopy or in stereoscopy<br />
as well, or with the assistance of an automatic image matching<br />
process.<br />
On another hand the measure of tie points is very automated: one can<br />
replace the precise measurement done by a human operator by automatic<br />
image-matching measures in very large number, either by the operator’s<br />
choice helped by the multi-windowing, or by automatic selection of tie<br />
points by an adequate algorithm. After filtering the wrong points, enough<br />
points normally remain to assure the stability of the block.<br />
Altimetry and digital terrain models (DTM)<br />
The DTM can be used either as supporting data to the restitution, or as<br />
a product. In addition, the digital photogrammetric workstations are<br />
perfectly adapted to the control and the correction of DTM obtained by<br />
automatic methods.<br />
In the case where a DTM pre-exists (result of a previous restitution, of<br />
a conversion of level lines . . .), it is possible to fix the measure mark to<br />
the DTM, which frees the operator from a tedious task. However, this<br />
help is quite limited, not only because of the frequently excessive generalization<br />
of the available models, and of the non-representation of the<br />
objects over the ground.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
<strong>Digital</strong> photogrammetric workstations 157<br />
In an intermediate way, it is possible to help the restitutor thanks to a<br />
real-time correlator, which replaces altimetric pointing. One uses in this<br />
case a correlator working in the object space, fixing the altitude to the<br />
maximum of the image matching detected on the vertical determined by<br />
the system of command. But the correlator must be considered as a help,<br />
and up to now it may work only in a supervised mode: many false correlations<br />
may happen, especially when the signal/noise ratio is low, or for<br />
example on specific periodic structures.<br />
The complete make-up of a DTM is usually let to a non-interactive<br />
process, possibly guided by some initial manual measures; indeed, the manual<br />
acquisition of a DTM is a long and very trying operation. One prefers<br />
to limit the human interventions to the control and the correction. The control<br />
is often performed visually, essentially by stereoscopic observation of<br />
level lines calculated from the DTM superimposed onto the couple of<br />
images. For corrections, each worker imagined some different solutions,<br />
from the local recalculation with different parameters, with possible integration<br />
of complementary human pointings, up to more brutal operations<br />
on altitudes themselves: new acquisition, levelling, digging, interpolation.<br />
Aids to the interpretation<br />
The functionalities of automatic or at least semi-automatic data extraction<br />
are awaited with impatience by users, who invested in systems of digital<br />
photogrammetry while hoping that algorithms of researchers would allow<br />
them to quickly improve their efficiency. But with the exception of automatic<br />
extraction of the DTM, and the cartographic restitution of contour lines, no<br />
tool is yet industrially implemented on systems of digital photogrammetry.<br />
Orthophotographies, perspective views<br />
<strong>Digital</strong> photogrammetric stations are often seen as machines to manufacture<br />
orthophotos. But actually, with the exception of the preparatory phases<br />
(reference points of the aerotriangulation, control and correction of the<br />
DTM) presented in the previous paragraph, it would be more economic<br />
to use a specialized software functioning on a standard computer.<br />
2.5.4.4 Limitations<br />
Photogrammetric workstations allow one, fundamentally, to perform at least<br />
the same work as analytic restitution devices, which they are progressively<br />
going to replace. Nevertheless, it is certain that the ocular comfort of operators<br />
is not equivalent to that of their predecessors, far from it. But as these<br />
materials directly follow the possibilities offered by the PC for video games,<br />
where the problem is quite similar, one can expect improvements of visual<br />
comfort, a point that employers are more and more obliged to take into<br />
account.
158 Raphaële Heno and Yves Egels<br />
In addition, these stations allow processes otherwise impossible to make<br />
on the same workstation, such as the realization of orthophotographies.<br />
Let us note finally that these stations obviously require images under<br />
digital shape, that are provided at present especially by digitization of<br />
argentic images, which implies a phase of quite expensive digitalization<br />
work currently, and a significant supplementary delay. The generalization<br />
of the use of digital cameras providing directly digital aerial images will<br />
make this situation evolve. In the intermediate situation that prevails to<br />
the date of writing of this work, it is essentially the price of acquisition<br />
and maintenance of equipment (that are hardly more powerful than simple<br />
PCs) and of software (that democratize themselves considerably) that make<br />
these stations attractive.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3 Generation of digital terrain<br />
and surface models<br />
INTRODUCTION<br />
The mathematic modelling of a landscape is a very important step of the<br />
photogrammetric processes. It has several key applications, and among<br />
them the cartography of the intervisibility (use of digital models for telecommunications,<br />
cellular phones for example), the hydrologic studies, the<br />
preparation of cartographic features (the contour lines), and the preparation<br />
of orthophotographies. Thus we shall review some important points<br />
that are not, strictly speaking, bound to the digital aspect of the photogrammetry<br />
itself, but that in fact are such a substantial output of today’s<br />
photogrammetry that it is necessary to analyse them here as completely<br />
as possible. We will see how are defined and specified the various surface<br />
models (§§3.1 and 3.2), how the data samples are produced (§3.3), how<br />
to remove raised structures when one wants to get a DTM (§3.4), how<br />
to build an optimal triangulation of a digital model so as to describe a<br />
surface in the best way possible (§3.5), how to extract automatically the<br />
characteristic terrain lines, as these lines are necessary to minimize<br />
unpleasant artefacts of any DTM, particularly for hydrologic matters<br />
(§3.6). Then we will conclude by some definitions, some quality considerations<br />
and some practical remarks about the production of digital<br />
orthophotographies (§§3.7, 3.8 and 3.9).<br />
3.1 OVERVIEW OF DIGITAL SURFACE MODELS<br />
Nicolas Paparoditis, Laurent Polidori<br />
3.1.1 DSM, DEM, DTM definitions<br />
A digital elevation model (DEM) is a digital and mathematical representation<br />
of an existing or virtual object and its environment, e.g. terrain<br />
undulations within a selected area. DEM is a generic concept that may<br />
refer to elevation of ground but also to any layer above the ground such
160 Nicolas Paparoditis and Laurent Polidori<br />
as canopy or buildings. When the information is limited to ground elevation,<br />
the DEM is called a digital terrain model (DTM) and provides<br />
information about the elevation of any point on ground or water surface.<br />
When the information contains the highest elevation of each point, coming<br />
from ground or above ground area, the DEM is called the digital surface<br />
model or DSM.<br />
Natural landscapes are too complex to be analytically modelled, so that<br />
the information is most often made of samples. Theoretically, a genuine<br />
‘model’ should also include an interpolation law that would give access<br />
to any elevation value between the samples, but this is generally left to<br />
the end user.<br />
Together with the elevation data, the specification of a DEM is provided<br />
by ancillary data. The specification, which is a description of the data set,<br />
is necessary to let users access, transmit or analyse the data. It consists of<br />
a number of characteristics that may be given as requirements.<br />
3.1.2 DEM specification<br />
The specification of a DEM generally includes two kinds of parameters.<br />
On the one hand standard altimetric specifications do not differ from<br />
the case of analogue maps, typically geodetic parameters (ellipsoid, projection,<br />
elevation origin . . .) and geographic location (e.g. coordinates of<br />
corners) but not scale, which is meaningless in the case of digital maps.<br />
On the other hand, a DEM is a digital product that cannot lead to an<br />
altimetric grid without a few specifications:<br />
• the digital format, i.e. a type (integer, character, real . . .) and length<br />
(often 2 bytes);<br />
• the significance of numerical values, i.e. unit (metre or foot) and in<br />
some cases, the coefficients of a conversion law, for instance a linear<br />
transform that makes the values fit a specified interval;<br />
• the grid structure, which may be irregular (e.g. triangular irregular<br />
networks or digitized contour lines) or regular (typically a square mesh<br />
regular grid);<br />
• the mesh size, which is important in the case of a square mesh regular<br />
grid – not to be considered as a resolution.<br />
The impact these specifications may have on the quality of the DEM is<br />
discussed in §3.2.<br />
The three main widespread models are regular raster grids, triangular<br />
or planar faces irregular networks, and iso-contour and break lines<br />
networks.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3.1.3 <strong>Digital</strong> models representation<br />
3.1.3.1 Regular raster grid (RG)<br />
Regular raster grids are well adapted to the representation of 2.5D surfaces,<br />
i.e. surfaces that can be described by a 3D mathematical function of the<br />
form z f(x, y). Regular raster grids describe a regularly sampled representation<br />
of this function f which implicitly assumes a definition of a<br />
projection (where the form z f(x, y) is valid) defining the expression of<br />
the (x, y) coordinates from the initial 3D coordinates in a global Cartesian<br />
reference system. When the surface describes the Earth’s relief this projection<br />
is a map projection given an ellipsoid.<br />
Indeed, RGs have the geometry of an image where the pixels are the<br />
nodes of the regular raster grid and the grey values of the pixels represent<br />
the elevations. Indeed, one of their main advantages is that they can<br />
be visualized as grey-level images or in colour with a look-up table, e.g.<br />
hypsometric. They should also preferably, for data size reasons, be stored<br />
as images. To pass from the image to the grid geometry and information,<br />
some parameters (x0 , y0 , dx, dy, b) and the units in which these parameters<br />
are expressed, in addition to the map projection system and to the<br />
ellipsoid parameters, have to be known and stored either in the image/grid<br />
header or aside in a separate file. The transformation from the image coordinates<br />
of pixel (i, j) to corresponding 3D coordinates (x, y, z) can be<br />
expressed as:<br />
x i dx x 0<br />
y j dy y 0<br />
Overview of digital surface models 161<br />
z G(i, j) dz b. (3.1)<br />
G(i, j) is the grey level of pixel (i, j). (x 0 , y 0 ) are the spatial coordinates of<br />
the image’s first row and line pixel. (dx, dy, dz) are the spatial sampling<br />
of the grid respectively along the x, y and z axes. b is the elevation corresponding<br />
to the grey-level 0 in the image. The dz and the b parameters<br />
are necessary if the grey levels are not coded as floating values but as 8<br />
or 16 bits integers.<br />
As all sampled models, a raster grid can describe many possible surfaces.<br />
This problem arises when we want to determine the elevation of a point<br />
falling inside the grid in between the known nodes of the grid. The elevation<br />
has to be calculated from the elevations of the neighbouring grid<br />
points with an interpolation function (bilinear, bicubic, etc.).<br />
The problem with the RG models is that the spatial sampling of the<br />
grid is regular and thus some features of the landscape can be correctly<br />
described at a given spatial sampling while some others, smaller, would<br />
not be relevantly sampled thus smoothed or even missing in the sampled<br />
model. Using the same density of samples all across a changing landscape
162 Nicolas Paparoditis and Laurent Polidori<br />
is definitely a limit of these models. Adaptive sampling meshes are more<br />
adapted to describe an irregular world. The distribution of elevation points<br />
needs to be dense on rough relief areas and only sparse on smooth areas.<br />
Indeed, the points do not need to be dense but to be well chosen to describe<br />
the surface as well as the application requires.<br />
Moreover, an underlying hypothesis made in aerial photogrammetry is<br />
that we suppose that the surface we are trying to model can be described<br />
by a graph of the form (x, y, f(x, y)). Indeed, this hypothesis is not always<br />
valid, e.g. in urban areas due to 3D discontinuities. It is even less valid in<br />
the case of terrestrial photogrammetry where overlapping surfaces often<br />
occur. RGs are an easy but not a general way to model surfaces.<br />
3.1.3.2 Triangular irregular networks (TIN)<br />
Most of the conventional data acquisition systems provide sparse point<br />
measurements. Building a regular grid DSM from these samples is thus often<br />
against nature. The idea of triangular irregular networks is to adapt completely<br />
the model to the samples by describing the surface by elementary<br />
triangles where the vertices are the samples themselves. These triangles can<br />
be constructed from the samples limited to their planimetric components<br />
by a 2D Delaunay triangulation process (a very good triangulation algorithm<br />
is available on the web site of the Carnegie Mellon University) if the surface<br />
is 2.5D and by a more complex 3D Delaunay triangulation process, also<br />
called tetraedrization, if not. This triangular modelling is widespread and<br />
extremely popular in virtual reality and in CAD world and systems.<br />
The triangles built by a triangulation process have only the aim of defining<br />
neighbourhoods in which one can directly calculate the elevation using<br />
an interpolation function between the three vertices for a given (X, Y).<br />
The simplest interpolation function is the following. Let T(P1 , P2 , P3 ) be<br />
the considered triangle where P1 (X1 , Y1, Z1 ), P2 (X2 , Y2, Z2 ), P3 (X3 , Y3, Z3 )<br />
are the three triangle vertices. Let V((X1 ,Y1 ,Z1 ) (X,Y,Z)), V1 {(X2 Y2 ,<br />
Z2 ) (X1 , Y1 , Z1 )), V2 ((X3 , Y3 , Z3 ) (X1 , Y1 , Z1 )), and (, , ) be the<br />
barycentric coordinates of point P inside T. V lies inside the plane defined<br />
by (V1 , V2 ) if V can be expressed in a unique way under the form V <br />
V1 V2 where:<br />
(V1∧V2 )k and v (V∧V1)k (V2∧V1 )k where k 0<br />
0<br />
1 .<br />
(V∧V 2)k<br />
(3.2)<br />
Thus Z (V 1 V 2 )k.<br />
Contrary to RGs, the samples and the set of triangles could be nonordered,<br />
thus leading to more time-consuming elementary operations as<br />
interpolating the Z elevation value for a given (X, Y). Indeed all triangles<br />
have to be parsed to determine the whole set of triangles including this
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
point. For a given (X, Y) and for each triangle T within the set of triangles<br />
we can calculate the coordinates of P(X, Y, Z) belonging to the plane lying<br />
on T as described above. Under the assumption that the surface is 2.5D,<br />
the parsing of the set of triangles can be stopped when P(X, Y, Z) belongs<br />
to T which is verified if and only if:<br />
<br />
1 0<br />
0<br />
0 .<br />
Overview of digital surface models 163<br />
(3.3)<br />
Adding some easy topology (ordering information) by storing for each<br />
triangle the index of the three adjacent triangles limits considerably the<br />
number of triangles to be parsed. Indeed, let us consider a triangle T 0 . We<br />
would like to determine the triangle T containing the point (X, Y). Starting<br />
from T 0, we will look for the adjacent triangle T 1 of T 0 in the direction<br />
of (X, Y). We will start this process again from T 1 and again until the<br />
current triangle contains (X, Y). This is also a way of parsing the triangles<br />
to generate an elevation profile between two 3D points on the surface.<br />
Some further topology that we will call spatial indexing can accelerate<br />
in an impressive way the number of triangles to test. Indeed, the 2D<br />
(x, y) space can be regularly and recursively split, e.g. in a dichotomy<br />
process, in square bounding boxes and the patches can be sorted in a tree<br />
graph where every branching node would represent a rectangular bounding<br />
box – which would get smaller while climbing up in the tree – and where<br />
the leaves at the end of the branches would be the patches themselves.<br />
One should remark that a patch/leaf could belong to several boxes/<br />
branches. We here obtain a hybrid mix of an irregular sampling with a<br />
regular ordering which keeps at the same time the useful flexibility of<br />
adaptive meshes to describe the relief variations, and the rapidity to access<br />
to interpolated information.<br />
This spatial indexing can also be given by an index raster map giving<br />
for each (X, Y) node of the map grid the index of the corresponding<br />
triangle. This map can be constructed by filling each triangle (limited to<br />
its planimetric components) with the label/index value of the triangle inside<br />
the map grid. The quality of this map will depend on its spatial sampling<br />
considering the aliasing problems arising close to the triangle limits.<br />
A drawback of raw TIN models is that the slope is identical on the<br />
whole facet surface and the slope is discontinuous between adjacent facets.<br />
If we suppose that our sample acquisition technique provides us in addition<br />
the surface normal vector and if we assume that the surface is smooth<br />
and curved inside the facet and/or across the facets, we can improve the<br />
surface approximation by adding to each triangle some more parameters<br />
to model locally the behaviour of the surface by an analytical function,<br />
e.g. bicubic splines, compatible in continuity and in derivability to all adjacent<br />
facets.
164 Nicolas Paparoditis and Laurent Polidori<br />
3.1.3.3 Surface characteristic lines<br />
A surface can also be described by characteristic feature lines (or points)<br />
and by contour lines (also called soft lines) giving the intersection between<br />
the surface and planes regularly sampled in one direction. This modelling<br />
is valid if the surface is 2.5D. Contour lines can be directly acquired from<br />
manual stereo-plotting of a stereopair, through analogue, analytic, or digital<br />
devices, derived from a regular grid or a TIN model DSM, or digitized<br />
manually or automatically from scanned maps. The reconstruction of a<br />
surface from contour lines is well conditioned if the surface is smooth. If<br />
not the addition of characteristic breaking lines (also called hard lines),<br />
e.g. slope break lines, helps in the regularization of the reconstruction<br />
problem by describing the surface local high derivatives.<br />
Modelling a surface in this way can also be seen as a data compression<br />
of the true surface. And the density of contour lines can be seen as a data<br />
compression rate vs. a data loss ratio.<br />
To conclude, the choice of one of these models and on its parameters<br />
will depend on the 3D geometric particularities of the object to model, on<br />
the requirements of the application using the model, and will have a definite<br />
impact on the surface approximation accuracy and on the data storage<br />
size.<br />
3.2 DSM QUALITY: INTERNAL AND EXTERNAL<br />
VALIDATION<br />
Laurent Polidori<br />
3.2.1 Basic comments on DSM quality<br />
Many different techniques can be used to extract a DSM, depending on<br />
available data, tools or know-how: map digitization and interpolation,<br />
optical image stereo correlation, shape from shading, interferometry, laser<br />
altimetry, ground survey, etc.<br />
In spite of their variety, each of these techniques can be described as a<br />
two-step process:<br />
• the first step consists in computing 3D locations for a large number<br />
of terrain points;<br />
• the second step consists in resampling the resulting data in order to<br />
fit a particular grid structure and a particular data format.<br />
Therefore, the quality of a DSM is the result of the way these two steps<br />
have been carried out.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
What is the quality of a digital surface model? Basically, it is its capability<br />
to describe the real surface, but quality criteria cannot be defined<br />
without keeping in mind user requirements. The impact of DSM quality<br />
on the reliability of derived information (such as geomorphology or hydrography)<br />
has been analysed by many authors to improve the understanding<br />
of what DSM quality means (e.g. Fisher, 1991; Lee et al., 1992; Polidori<br />
and Chorowicz, 1993).<br />
The relevance of a criterion depends on the way it reflects these requirements,<br />
and on the feasibility of an efficient validation to check its fulfilment.<br />
3.2.2 Quality factors for DSM<br />
DSM quality 165<br />
As recalled above, the quality of a digital surface model is affected by both<br />
point location accuracy and resampling efficiency.<br />
Point location accuracy<br />
Each point location technique has a specific error budget, with contributions<br />
from the intrinsic features of the employed system (sensor, platform,<br />
etc.) and from the acquisition and processing parameters.<br />
Most DSM extraction techniques (and particularly photogrammetric and<br />
profiling techniques) are direct techniques, which means that they provide<br />
point location measurements in which adjacent samples are independent<br />
from each other. They are more suitable for elevation mapping than for<br />
slope mapping. On the contrary, differential techniques (such as shape<br />
from shading or radar interferometry) provide information about surface<br />
orientation, i.e. slope or azimuth. In this case, elevations are obtained by<br />
slope integration, so that height errors are subject to propagation. This<br />
phenomenon may be reduced by crossing different integration paths.<br />
As far as resampling is concerned, its impact on DSM quality depends<br />
on grid geometry (structure and density) and on data format.<br />
Grid structure<br />
A great variety of grid structures has been proposed for digital surface<br />
model resampling. They have been reviewed and discussed by several<br />
authors (Burrough, 1986; Carter, 1988). Three main approaches can be<br />
mentioned:<br />
• regular sampling, in which all meshes have constant size and shape<br />
(most often square);<br />
• semi-regular sampling, which is based on a very dense regular grid in<br />
which only useful points have been selected;<br />
• irregular sampling, where terrain points may be located anywhere.
166 Laurent Polidori<br />
Regular sampling has obvious advantages in terms of storage, but it is not<br />
very efficient to depict natural relief in which shapes and textures are mainly<br />
irregular, unless the grid is considerably densified. This is the reason why<br />
semi-regular methods like progressive or composite sampling (Burrough,<br />
1986; Charif and Makarovic, 1989), and irregular ones like TINs<br />
(Burrough, 1986; Chen and Guevara, 1987) are so successful for digital<br />
terrain modelling.<br />
Grid density<br />
The density of a DSM is basically a trade-off between economical<br />
constraints (which in general tend to limit the density) and accuracy requirements<br />
(which are rather fulfilled by a higher grid density).<br />
The impact of mesh size on DSM quality has mainly been studied in<br />
terms of height error (e.g. by Li, 1992), but its major effect is on height<br />
derivatives. Indeed, reducing the density of a DSM grid (i.e. subsampling)<br />
removes the steepest slopes and makes the surface model smoother.<br />
Data format<br />
The digital format of the data has to be mentioned as well within the<br />
contributors to DSM quality, and in particular the number of bytes per<br />
sample. Formats using 2 bytes (i.e. 65,536 levels) which allow a 10 cm<br />
precision over an elevation range of more than 6,000 m, are commonly<br />
used because they provide a good trade-off between 1 byte (which limits<br />
the accuracy) and 4 bytes (which increases the volume of data needlessly).<br />
3.2.3 Internal validation<br />
A set of 3D coordinates drawn at random has very little chance of yielding<br />
a realistic relief. Therefore, it is worth checking that the terrain described<br />
by the DSM is possible, i.e. that it fulfils the main properties of real topographic<br />
surfaces. These properties can be quite straightforward. For<br />
instance ‘rivers go down’ or, in urban areas, ‘building walls are vertical’.<br />
Checking to what extent these properties are respected does not require<br />
reference data but only a generic knowledge of landscape features. For<br />
this reason, it can be called internal validation.<br />
Visual artefact detection, which should always be done before DSM<br />
delivery, is the first level of internal validation.<br />
Unrealistic textures, such as strips or other anisotropic features, can be<br />
revealed by Fourier analysis or by comparing variograms in different directions.<br />
These artefacts can result from image scanning (Brown and Bara,<br />
1994) or from contour line interpolation (Polidori et al., 1991).<br />
If some rules are universal (e.g. rivers go down), others require an expert<br />
knowledge about geomorphology (for natural relief mapping) or about
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
urban structure design (for building extraction). For instance, texture<br />
isotropy is not supposed to be fulfilled everywhere, so that some subjectivity<br />
is required to distinguish natural anisotropies from directional<br />
artefacts. This distinction will be easier if the DSM extraction algorithms<br />
are known, because each extraction step may have a contribution to the<br />
observed artefacts: correlation noise, original image striping, grid resampling,<br />
etc.<br />
3.2.4 External validation<br />
If external elevation data are available and reliable, an external validation<br />
can be considered, which consists in comparing the DSM with the reference<br />
data. This is the most usual way of evaluating the quality of DSMs,<br />
but it is limited by two major difficulties.<br />
The first difficulty is the availability of a suitable reference data set.<br />
Indeed, DSMs are often validated with very few ground control points, so<br />
that the comparison may be statistically meaningless. Moreover, these<br />
control points have their own error which in most cases is not known and<br />
which may have the same order of magnitude as the DSM they are supposed<br />
to control.<br />
The second difficulty is the need for an explicit comparison criterion,<br />
which must reflect application requirements. On the one hand, the magnitude<br />
to be compared has to be defined: it is altitude in most cases, but<br />
slope or other derivative magnitudes could also be considered. On the<br />
other hand, a statistical index has to be defined too, generally based on<br />
the histogram of height differences: mean, standard deviation, maximum<br />
error, etc. have different meanings.<br />
An interesting way of overcoming these difficulties is to map the discrepancies<br />
in order to display their spatial behaviour. This validation approach<br />
is not quantitative, but it proceeds very useful information, such as which<br />
landscapes (in terms of relief shape but also land cover) are accurately<br />
depicted and which are poorly shown. Therefore, it may contribute to<br />
improve the understanding of a given DSM extraction technique.<br />
References<br />
DSM quality 167<br />
Brown D., Bara T. (1994) Recognition and reduction of systematic error in elevation<br />
and derivative surfaces from 7 1/2 minute DEMs. Photogrammetric<br />
Engineering and Remote Sensing, vol. 60, no. 2, pp. 189–194.<br />
Burrough P. (1986) Principal of geographic information systems for land resources<br />
assessment. Oxford University Press, New York.<br />
Carter J. (1988) <strong>Digital</strong> representations of topographic surfaces. Photogrammetric<br />
Engineering and Remote Sensing, vol. 54, no. 11, pp. 1577–1580.<br />
Charif M., Makarovic B. (1989) Optimizing progressive and composite sampling<br />
for DTMs. ITC Journal, vol. 2, pp. 104–111.
168 Laurent Polidori<br />
Chen C., Guevara J.A. (1987) Systematic selection of very important points (VIP)<br />
from digital elevation model for constructing triangular irregular networks.<br />
Proceedings of Auto Carto 8, Baltimore, 29 March–3 April, pp. 50–56.<br />
Fisher P. (1991) First experiments in viewshed uncertainty: the accuracy of the<br />
viewshed area. Photogrammetric Engineering and Remote Sensing, vol. 57, no.<br />
10, pp. 1321–1327.<br />
Lee J., Snider P., Fisher P. (1992) Modelling the effect of data error on feature<br />
extraction from digital elevation models. Photogrammetric Engineering and<br />
Remote Sensing, vol. 58, no. 10, pp. 1461–1467.<br />
Li Z. (1992) Variation of the accuracy of digital terrain models with sampling<br />
interval. Photogrammetric Record, vol. 14(79), pp. 113–128.<br />
Polidori L., Chorowicz J., Guillande R. (1991) Description of terrain as a fractal<br />
surface and application to digital elevation model quality assessment. Photogrammetric<br />
Engineering and Remote Sensing, vol. 57, no. 10, pp. 1329–1332.<br />
Polidori L., Chorowicz J. (1993) Comparison of bilinear and Brownian interpolation<br />
for digital elevation models. ISPRS Journal of <strong>Photogrammetry</strong> and<br />
Remote Sensing, vol. 48(2), pp. 18–23.<br />
3.3 3D DATA ACQUISITION FROM VISIBLE IMAGES<br />
Nicolas Paparoditis, Olivier Dissard<br />
Introduction<br />
The surface that can be processed from images is an observable surface.<br />
In the case of aerial visible images, this surface will describe the terrain<br />
but also all the structures lying on the terrain at the scale/resolution of<br />
the images e.g. buildings, vegetation canopy, bridges, etc.<br />
The techniques presented here are applicable to visible images of all<br />
platforms (helicopters, airplanes, satellites, ground vehicle, ground static,<br />
etc.) and of all kinds of passive sensors (photos, CCD pushbroom, CCD<br />
frame cameras) as long as the internal, relative and external parameters<br />
for each view are known. In other words, we suppose that the platformtriangulation<br />
process has already been achieved previously. Nevertheless,<br />
even though the concepts remain the same, the 3D processes and the integration<br />
of the processes themselves can change slightly, depending on the<br />
survey acquisition possibilities and constraints. We will point out whenever<br />
the particularities of the sensor and the platform have an impact on<br />
the processes described.<br />
Besides the particular process assessment induced by the particular geometry<br />
and distribution of the data, the methods and the strategies to be<br />
involved also depend on what kind of 3D data we want to derive from<br />
the visible images, i.e. single-point measurements or more dense and more<br />
regular for the generation of a DEM. We will start from the simplest, the<br />
stereopair processing of an individual 3D measurement, before looking at
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
the global processing for the generation of an elementary stereopair DEM<br />
and a global DEM over the survey, and we will end by defining the proper<br />
specifications for a survey acquisition system (and the processes to manage<br />
them) enabling the processes to avoid in a clean way all (or at least most)<br />
of the problems and allowing us to reach the automation of a complete<br />
and robust enough DEM on the whole survey.<br />
3.3.1 Formulation of an automatic individual point<br />
matching process<br />
Nicolas Paparoditis<br />
The major interest of digital photogrammetry, besides all the management<br />
facilities that it provides, is that images are described by a set of regularly<br />
spaced digital values that can be easily manipulated with algorithms. Let<br />
I 1 and I 2 be the two digital images composing the stereopair, (i, j) a point<br />
position in one of the image grids, and R(i, j) the corresponding grey-level<br />
value.<br />
The concern of digital matching, is how do we determine that any point<br />
(i 1, j 1) of I 1 and point (i 2, j 2) of I 2 are alike or homologous? And more generally,<br />
how do we determine that characteristic features of I 1 and I 2 are alike? This<br />
preoccupation is one of the most recurrent in digital photogrammetry, image<br />
processing and in automatic pattern recognition techniques.<br />
Entity attributes and similarity measure<br />
As we are looking for individual point matches, we will stick for now with<br />
point entities. Most of the time a function f derives from image measurements<br />
(values) on or around our point entities (i1, j1) and (i2, j2) corresponding<br />
entity attribute vectors V1 and V2 . Then a similarity function g<br />
(g can be a distance but does not have to be) derives a numerical similarity<br />
value (or score) C describing the similarity between V1 and V2 and<br />
thus the entities themselves. C can be expressed as follows:<br />
C((i 1, j 1), (i 2, j 2)) g(V 1, V 2).<br />
3D data acquisition from visible images 169<br />
Interpolation function<br />
We point out that the image sampling effects are such that the real homologue<br />
(i2, j2) of a point (i1, j1) on the grid of image 1 is not itself on the<br />
grid in image 2. Thus the function f should implicitly use and be built on<br />
an interpolation function V (bilinear, bicubic, sinc, etc.) which determines<br />
the intensity value in any sub-pixel point of an image from its closest<br />
neighbours on the image grid.
170 Nicolas Paparoditis and Olivier Dissard<br />
An ill-posed and a combinatorial problem<br />
In practice, how do we build V and choose f ? Let us take as an example<br />
the simplest f and g. For instance, f giving the intensity value itself and g<br />
the difference between the intensity values. Due to the image noise, the<br />
radiometric quantification, and the image sampling, the grey-level values<br />
themselves for real homologous points are in general different. Thus, many<br />
points in I1 will have close intensity values and for a given point in I1 many points in I2 will have close intensity values. Thus we have a serious<br />
matching ambiguity problem and we can conclude that the information<br />
derived from f is not characteristic enough and g not stable and discriminating<br />
enough. We also face a combinatorial explosion especially if we<br />
want to match all points of I1. The problem of digital image matching belongs to the family of mathematically<br />
ill-posed problems (Bertero et al., 1988). The existence, the<br />
uniqueness, the stability of a solution are not a priori guaranteed. This<br />
problem can be transformed in a well-posed one by imposing regularizing<br />
constraints to diminish the degrees of freedom in the parameter space so<br />
that the domain of possible solutions can be reduced (Tikhonov, 1963).<br />
Area-based matching: a first solution to ambiguity<br />
One solution to this matching ambiguity problem is to make the entities<br />
attribute vectors as unique as possible. Instead of matching the grey level<br />
of the pixel, we can match the context (all the grey levels) around the<br />
homologous points. We here make the assumption that the contexts are<br />
also homologous. The context is most of the time characterized by all the<br />
grey levels of the pixels inside a rectangular (most of the time square)<br />
template window (also called image patch) centred on the entity thus with<br />
34<br />
46<br />
54<br />
57<br />
75<br />
90<br />
93<br />
95<br />
56<br />
67<br />
65<br />
60<br />
69<br />
73<br />
80<br />
76<br />
134<br />
67<br />
66<br />
63<br />
75<br />
85<br />
76<br />
94<br />
Figure 3.3.1<br />
100<br />
87<br />
54<br />
54<br />
51<br />
65<br />
48<br />
108<br />
103<br />
94<br />
75<br />
60<br />
54<br />
58<br />
54<br />
102<br />
105<br />
100<br />
87<br />
76<br />
56<br />
57<br />
75<br />
101<br />
120<br />
110<br />
91<br />
80<br />
67<br />
63<br />
86<br />
103<br />
107<br />
109<br />
93<br />
80<br />
73<br />
87<br />
89<br />
101<br />
(i, j)<br />
⎛54<br />
⎞<br />
⎜ ⎟<br />
⎜75<br />
⎟<br />
⎜87<br />
⎟<br />
⎜ ⎟<br />
⎜54<br />
⎟<br />
⎜<br />
v(i, j) = 60<br />
⎟<br />
⎜ ⎟<br />
⎜76<br />
⎟<br />
⎜ ⎟<br />
⎜51<br />
⎟<br />
⎜54<br />
⎟<br />
⎜ ⎟<br />
⎝56<br />
⎠
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
an odd number of lines N and rows M. The entity can be described by<br />
a vector v with N × M components. Each component is a grey level of a<br />
template pixel following a given spatial ordering (as shown in Figure 3.3.1).<br />
Thus, the vector describes the local image texture pattern, which is of<br />
course much more discriminating than a single intensity value. The larger<br />
the templates, the higher the discrimination power. Meanwhile, our context<br />
stability assumption becomes less and less valid (see ‘Geometric assumptions’<br />
hereunder).<br />
Matching the entities is equivalent to matching their area context vectors.<br />
Matching these vectors is what we call area-based matching. The requirement<br />
for area similarity functions is that their response is optimal for real<br />
homologous points and that they are as stable as possible to some radiometric<br />
and geometric changes between homologous template areas in both<br />
images. The most commonly used similarity functions on these areas are<br />
the following:<br />
Least squares differences:<br />
C1((i1, j1), (i2, j2)) <br />
Scalar product:<br />
|| ; (3.4)<br />
V2(i2, j2) V2(i2, j2) V 2<br />
1(i1, j1) V1(i1, j1) ||<br />
||<br />
3D data acquisition from visible images 171<br />
C2((i1, j1), (i2, j2)) . (3.5)<br />
V 2<br />
1(i1, j1)V2(i2, j2) V 1 (i 1 , j 1 ) V 2 (i 2 , j 2 ) ||<br />
As shown in Figure 3.3.2, these similarity functions have a geometrical<br />
explanation. C 1 is the norm of V 3 and C 2 is cos , where is the angle<br />
between V 1 and V 2. When the textures are alike V 1 and V 2 should be<br />
nearly collinear. Thus C 1 should be close to 0 and C 2 close to 1.<br />
If corresponding templates have the same texture but are affected by a<br />
radiometric shift (case where V 1 ~ V 2 ), due to the fact that the images<br />
could be acquired with different lighting conditions or in the case of<br />
0<br />
v 1<br />
Figure 3.3.2<br />
θ<br />
v 3<br />
v 2
172 Nicolas Paparoditis and Olivier Dissard<br />
analogue photographs that the images have been scanned in different conditions<br />
or that the observed surface has a non-Lambertian behaviour, the<br />
similarity scores will decrease significantly thus altering the discrimination<br />
process. To overcome this problem the vectors can be centred. Thus if 1<br />
(resp. 2 ) is the mean grey level and 1 (resp. 2 ) is the root mean square<br />
of grey levels of the template around (i 1 , j 1 ) (resp. (i 2 , j 2 )) the previous similarity<br />
functions become:<br />
Centred least squares:<br />
C′ 1((i1, j1), (i2, j2)) || ; (3.6)<br />
V2(i2, j2) 1 V 2<br />
1(i1, j1) 2 ||<br />
Linear cross-correlation:<br />
V<br />
C′ 2((i1, j1), (i2, j2)) <br />
1(i1, j1)V2(i2, j2) 12 . (3.7)<br />
Could not the images be pre-processed with a global histogram radiometric<br />
equalization instead of centring each vector? Indeed not! This<br />
process would alter the image quality and consequently the quality of the<br />
matching process. Furthermore, some of these effects change through the<br />
images, i.e. hot spot and non-Lambertian effects. Could the contrasts be<br />
enhanced? No! In general we do so when we look at an image, because<br />
of our non-linear and low differential eye sensibility. Nevertheless, a correlation<br />
process has no such problems.<br />
Image matching definition<br />
Under the assumption the matching problem is well posed, if (i2, j2) is the<br />
homologue of (i1, j1) inside the admissible domain of hypotheses S within<br />
I2 then one of the two following expressions should be verified for a given<br />
similarity function C:<br />
or<br />
(i 2, j 2) <br />
ArgMax<br />
(i, j)∈I 2 /S<br />
(i2, j2) ArgMin C V1(i1, j1), V2(i, j) . (3.8)<br />
(i, j)∈I 2 /S<br />
1<br />
1 2<br />
C V 1(i 1, j 1), V 2(i, j)<br />
Geometric assumptions<br />
The area-based matching process described above rests upon a major<br />
assumption. Indeed, we have implicitly supposed that the homologous<br />
2
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 173<br />
template of image 1 can be found in image 2 by a simple shift. In general,<br />
the differences in the viewing geometry (due to erratic variations of attitude<br />
of the sensor) between the two views and the deformations due to the<br />
landscape relief are such that this assumption is not strictly verified.<br />
Feature-based matching<br />
If the point entity itself is or belongs to a characteristic image feature, e.g.<br />
an interest point (see §2.4) or a contour point, the homologous point can<br />
be looked for in the set of identical image feature entities in the other<br />
image. This will restrict the combinatorial problem and thus the ambiguity<br />
problem. Furthermore, these entities are less sensitive to geometric<br />
deformations between the images. Information such as local gradient norm<br />
or direction, second or higher degree image derivatives can be used to<br />
build the feature entity attribute vectors. Nevertheless, in an urban landscape<br />
for instance, many image features look alike and finding characteristic<br />
and discriminating entity attributes is difficult and sometimes impossible.<br />
Moreover, the entities themselves, e.g. building contours, are not stable<br />
thus the ambiguity problem remains. Feature-based matching is also an<br />
ill-posed problem and is limited to a sparse set of points in the scene and<br />
thus does not provide the universality of area-based techniques for matching<br />
every point in the scene or in the images.<br />
Stereopair geometric constraints: a second solution to the<br />
ambiguity problem<br />
Another way of limiting the ambiguity problem is by reducing the combinatory<br />
of the matching problem, i.e. the search space for homologous<br />
points. Indeed, the bigger the search space the higher the number of possible<br />
homologous hypotheses and the higher the probability of encountering<br />
matching ambiguities.<br />
The following paragraphs describes the different matching methods that<br />
can be applied to an oriented stereopair to overcome the geometric distortion<br />
problems and to restrict the spatial domain S of homologous solutions.<br />
Generally the point-to-point matching problem in a stereopair geometry<br />
can be expressed in two radically different ways and thus leads to two<br />
different matching methods each having different advantages.<br />
3.3.2 Stereo matching from image space (SMI)<br />
Nicolas Paparoditis<br />
This first method only requires the relative orientation of the images for<br />
the matching process. If the real 3D localization is meant to be computed,<br />
the absolute orientation of the images are required. The matching
174 Nicolas Paparoditis and Olivier Dissard<br />
Figure 3.3.3<br />
P 1 (i 1, j 1)<br />
C 1<br />
O 1<br />
formulation of this method, which is image-based, can be expressed in the<br />
following terms: ‘given a (i 1, j 1) in I 1 (called the master or reference image)<br />
what is the corresponding (i 2, j 2) in I 2 (called the slave or secondary image)?’<br />
Epipolar lines<br />
The position of (i2 , j2 ) is geometrically (stereoscopically) constrained by<br />
(i1, j1). As we can see in Figure 3.3.3, all possible matches are on a line<br />
denoted E in the image 2, called the epipolar line, which is the projection<br />
in the image plane of the 3D plane lying on the 3D ray of (i1 , j1 ) including<br />
C1 and on C2 . Due to geometrical distortions of the optics, this line is<br />
more likely to be a curve. Nevertheless, the general 2D matching problem<br />
is now transformed into an easier 1D matching problem where (i2 , j2 ) has<br />
Figure 3.3.4<br />
C 1<br />
C 2<br />
E<br />
O 2<br />
C 2<br />
Motion
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 3.3.5<br />
O′ 1<br />
C 1<br />
P 1(i, j)<br />
O 1<br />
y<br />
3D data acquisition from visible images 175<br />
P′ 1(i′, j′)<br />
z<br />
to be searched for along this curve. Note that all epipolar lines corresponding<br />
to all points of image 1 intersect themselves in a point of image<br />
2, called epipolar point, which is the projection of C 1 in the image 2 plane<br />
(see Figure 3.3.4).<br />
If I 1 and I 2 are two consecutive images in the same aerial flight strip<br />
(near to vertical viewing axis), or classical across track (up to 50° of latitude<br />
so that we can consider the orbits to be locally parallel), or along<br />
track satellite stereopairs, homologous texture vectors of a given landscape<br />
patch (with a low slope) can be found by a simple translation. Thus our<br />
previous correlation scores similarity criteria can be applied directly. In<br />
these cases, the epipolar lines are very close to the image lines themselves.<br />
Nevertheless, for some aerial surveys, the differential yaw – but also<br />
pitch and roll – angles between both views, due to plane instability, can<br />
sometimes occur. The higher the difference is and the larger the window<br />
size gets, the less the texture vectors for homologous points will look alike,<br />
and the more the automatic matching process is unlikely to give good<br />
results. In these conditions to overcome this problem, a gyro-stabilized<br />
platform can be used so that epipolar lines follow as closely as possible<br />
the image lines (see Figure 3.3.5).<br />
Epipolar resampling<br />
To ease the algorithmic implementation of the matching process, the images<br />
are often resampled in a way that the epipolar lines become parallel and<br />
aligned on the image lines. Let I′ 1 and I′ 2 be the resampled images. Thus<br />
looking for the homologous estimate of point (i1, j1) can now be expressed:<br />
x<br />
O′ 2<br />
C 2<br />
O 2
(î 2, j , (3.9)<br />
where N is the number of rows of I′ 2 .<br />
How can we manage this resampling? Let (O1 , i1 , j1 , k1 ) and (O2 , i2 , j2 ,<br />
k2) be the two reference systems describing the orientation of the image<br />
planes. We simulate for each view the image of a virtual sensor (I′ 1 and<br />
I′ 2) with the same optical centres but with new reference systems and<br />
without distortions of the optics and of the focal plane. The new reference<br />
systems (O1, i′ 1, j′ 1, k′ 1) and (O2, i′ 2, j′ 2, k′ 2) are constructed so that i′ 1<br />
and i′ 2 are collinear to C1C2 , k′ 1 and k′ 2 are collinear to z, and j′ 1 k′ 1 ∧<br />
i′ 1 and j′ 2 k′ 2 ∧ i′ 2 .<br />
How do we construct the pixel grey levels of these new images? For<br />
each pixel P′(i′, j′) of image I′, we determine with our new sensor’s geometry<br />
the corresponding ray in 3D taking into account the distortion of<br />
the camera. According to previous geometry, we determine the corresponding<br />
image position P(i, j) in image I by intersecting the 3D ray with<br />
its image plane. Of course, this point falls anywhere inside the image grid,<br />
thus the grey-level value for P′(i′, j′) has to be calculated by interpolation<br />
of the grey-level values of neighbour nodes of P(i, j) inside the image grid.<br />
If it can be avoided, in the case of classical good conditions surveys, it<br />
is preferable. Indeed, this operation leads to a degradation of image quality<br />
which will alter slightly the quality of the image matching process. The<br />
only advantage of a systematic epipolar resampling is that the images can<br />
be viewed directly in digital photogrammetric workstations whatever the<br />
differential yaw between the images. (See Figure 3.3.6.)<br />
ˆ 176 Nicolas Paparoditis and Olivier Dissard<br />
2) (î 2, j1) ArgOpt C V (i I′1 1, j1), V (x, j I′2 1)<br />
x∈[0, N]<br />
(a) (b)<br />
Figure 3.3.6 Epipolar matching. The left image (a) and the right image<br />
(b) extracts are the result of the stereopair resampling. The broken lines<br />
correspond to the horizontal conjuguate epipolar lines in these resampled<br />
images.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
P 1 (i 1, j 1)<br />
C 1<br />
O 1<br />
Figure 3.3.7 Defining a search space interval.<br />
3D data acquisition from visible images 177<br />
O2 E′ (imin, (i j<br />
min 1) ,j1) (imax, j1) Defining an interval for the match search<br />
The range of positions for all possible homologous hypotheses for a given<br />
P1(i1, j1) in I1 can be delimited inside an interval ([i2min, i2max], j1) on the<br />
epipolar E′ by the knowledge of minimal and maximal z values (or the range<br />
of disparities/parallax if we only have the relative orientations) on the scene<br />
(as shown on Figure 3.3.7) if we want to have a given disparity range for<br />
all the points in image I1, or a gross relief model if we want to have this<br />
information more locally. This information could be provided by existing<br />
databases, e.g. world-covering low-resolution DTM or by a map.<br />
Another way of doing this, is to stereoplot manually the homologous<br />
image points for the two landscape points corresponding to the lowest and<br />
the highest point in the scene or plotting a larger set of points regularly<br />
sampling the scene if we want to have a locally adaptive range definition.<br />
Some more elaborate ways of reducing these intervals will be given in<br />
§3.3.8.<br />
Search-space sampling and homologous position estimation<br />
If di is the spatial sampling in I2 along the search space interval [i2min ,<br />
i2max], all the correlation scores for all the possible moves form a graph<br />
usually called a correlation profile. The homologous point is considered<br />
to be the point where the maximum of the similarity score (also called the<br />
correlation peak) occurs (see Figure 3.3.8). Its position can be expressed<br />
as follows:<br />
C 2<br />
Z min<br />
Z max
178 Nicolas Paparoditis and Olivier Dissard<br />
C<br />
i min<br />
d i<br />
î 2<br />
î 2subpix<br />
î 2 ArgMax C2V′ 1(i1, j1), V′ 2(imink·di, j1)imin) . (3.10)<br />
k∈[0,(imaximin)/di] The smaller the di the better the correlation profile is sampled and the<br />
higher the matching precision. In practice, to limit the number of similarity<br />
scores calculation, most of the time di is fixed to 1 so that all the<br />
homologous points hypotheses fall on the grid and thus the texture vectors<br />
are directly extractable from the images without any further interpolation<br />
processing. If we consider the peak to be the integer position of the homologous<br />
point, the maximum error in the homologous estimation is 0.5 pixels.<br />
A more precise (in most cases) sub-pixel position can be found by fitting<br />
a polynomial function, often a parabola, through the correlation values<br />
around the peak position. The most commonly used fitting function is the<br />
parabola. With the parabola, the sub-pixel position of the correlation peak<br />
can be obtained from the discrete positions as follows:<br />
C(î<br />
î2Subpix î2 <br />
21, j1) C(î21, j1) . (3.11)<br />
2C(î21, j1 ) C(î21, j1 ) 2C(î2 , j1 )<br />
If the di sampling is sub-pixel, with the idea of finding directly a correlation<br />
peak closer to the real homologous position without any a posteriori<br />
interpolation, then the homologous hypotheses samples fall inside the grid<br />
and thus all the components of the texture vectors have to be interpolated.<br />
Generally a di of one-fifth of a pixel reaches the limit of the matching<br />
process precision. One should remark that the grey levels of texture vectors<br />
of image I2 will have gone through two interpolation processes. We have<br />
pointed out earlier that a resampling alters the image quality and consequently<br />
the matching quality. In this particular case, this direct sub-pixel<br />
matching process should be carried out on the non-resampled images along<br />
the epipolar curve E.<br />
i max<br />
Figure 3.3.8 Sub-pixel localization of the correlation peak.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 179<br />
Doing the interpolation on the images or on the correlation profile,<br />
provides very close results. If the matching process has to be carried out<br />
on many points of image I 1 , one would prefer using the first technique as<br />
the number of samples to be calculated is smaller.<br />
Managing with relative orientation errors<br />
Due to the relative orientation errors, the real homologue of (i1, j1) is not<br />
on the computed epipolar line E′ itself but in a close range of the epipolar<br />
line. To cope with that the search space can be extended to all points of<br />
image I2 in a neighbourhood (depending on relative orientation residues;<br />
1 pixel is usually enough) on each side of the epipolar line. Thus our correlation<br />
profile becomes a correlation surface. If the peak of the correlation<br />
surface occurs for a point on the epipolar line, the easiest way of finding<br />
the sub-pixel position is to first determine the sub-pixel position in the<br />
epipolar direction by fitting a parabola through the correlation scores and<br />
then finding the sub-pixel position in the orthogonal direction by fitting<br />
another parabola. If not the neighbourhood should first be extended before<br />
a further process.<br />
3D localization of homologous points<br />
The 3D localization of homologous points (i1, j1) and (i2, j2) is obtained<br />
by intersecting both corresponding 3D rays D1(C1, R1) and D2(C2, R2). If<br />
(i2, j2) does not lie exactly on the epipolar line of (i1, j1), those rays do not<br />
intersect, as shown on Figure 3.3.9.<br />
P 1 (i 1, j 1)<br />
C 1<br />
O 1<br />
R 1 D1<br />
M 1<br />
P(x, y, z)<br />
D 2<br />
M2 Relative error<br />
Figure 3.3.9 3D triangulation and localization of homologous points.<br />
R 2<br />
O 2<br />
C 2<br />
P 2 (i 2, j 2)
180 Nicolas Paparoditis and Olivier Dissard<br />
Let M 1 be the closest point of ray D 1 to ray D 2 and M 2 the closest point<br />
of ray D 2 to ray D 1. M 1 and M 2 are such that:<br />
⎧<br />
⎪<br />
⎨<br />
⎪<br />
⎩<br />
M 1 C 1 1 R 1<br />
M 2 C 2 2 R 2<br />
M 1M 2 · R 1 0<br />
M 1M 2 · R 2 0<br />
⎧<br />
⎪<br />
thus ⎨<br />
(3.12)<br />
⎪<br />
⎩ 2 (C1C2, R1, R1∧R2) (R1∧R2) 2 (R1∧R2) .<br />
2<br />
The corresponding 3D point is generally chosen to be P (M 1 M 2 )/2,<br />
the closest and equidistant point to both rays. The ||M 1M 2|| distance gives<br />
us a quality estimator of the 3D relative localization.<br />
Relation between 3D precision and matching precision<br />
The 3D localization precision is affected by the aerial triangulation errors,<br />
the matching errors and its planimetric and altimetric components depend<br />
on the viewing angles and on the stereoscopic base to height ratio (B/H)<br />
as shown on Figure 3.3.10. Putting aside the errors due to the aerial triangulation<br />
process, thus making the assumption that the relative orientation<br />
is perfect, we can estimate theoretically the intrinsic precision of the<br />
matching process. Let ematch (resp. ealti) be the matching (resp. altimetric)<br />
error and match (resp. alti ) the root mean square of the matching (resp.<br />
altimetric) error expressed in pixels and r0 the ground pixel size in metres.<br />
Thanks to the Thales theorem, the altimetric error is given by:<br />
ealti H<br />
B r0 ematch and H<br />
alti <br />
B r0match .<br />
1 (C 1C 2, R 2, R 1∧R 2)<br />
The planimetric error depends on the position of (i 1, j 1) in image I 1:<br />
eplani tan (i) · ealti OP1 ,<br />
where f is the focal length expressed in pixels.<br />
The formulas that we have expressed here are not the models of the 3D<br />
errors that we would have if we intended plotting a given ground feature<br />
in image I1 and finding automatically the 3D corresponding point. Indeed<br />
the matched point (i2, j2) will correspond to (i1, j1) and not to the feature<br />
that was meant to be plotted in image I1. This plotting error will inevitably<br />
lead to another source of error in the 3D localization. Only the plotting<br />
errors along the epipolar lines will influence the altimetric precision. Let<br />
the parallax be p = O2P2 O1P1 and its errors along the epipolar lines<br />
dp dp.i ematch eplottingi then the errors are given by:<br />
ealti f
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
P 1(i 1, j 1)<br />
C 1<br />
Figure 3.3.10<br />
i<br />
O 1<br />
f<br />
3D data acquisition from visible images 181<br />
e alti<br />
ealti H<br />
B r0 dp and H<br />
alti <br />
B r0 √ 2 match 2 plotting<br />
O2 ematch P 2(i 2, j 2)<br />
. (3.13)<br />
The higher the base to height ratio, the better the altimetric precision.<br />
The choice of this ratio for a survey is conditioned by the desired quality<br />
for the mapping output. For a DEM production line, for instance, a higher<br />
one should preferably be chosen. Against that, when the ratio increases,<br />
the image differences and consequently the matching difficulty increases.<br />
(See Figure 3.3.11.)<br />
Intrinsic matching precision<br />
What is the precision that we can achieve from digital automatic matching?<br />
The limits of the matching precision are easy to estimate in the best case,<br />
i.e. when the homologous pattern can be found by a simple translation<br />
(horizontal images planes, flat landscape). To generate a virtual global<br />
translation between I1 and I2 , we can simulate a horizontal image plane<br />
stereopair with perfectly known orientation parameters on a perfectly flat<br />
landscape where we can map any kind of image pattern, e.g. an aerial<br />
orthophoto. Since we have simulated the images, we know exactly the<br />
translation value between I1 and I2. If we apply our matching process to<br />
estimate these translations on different stereopairs simulated for different<br />
translation values (fractional part between 0 and 1), we find systematic<br />
errors in the translation estimation depending on the real translation fractional<br />
part as shown in Figure 3.3.12.<br />
B<br />
r 0.e match<br />
e plani<br />
E′<br />
P<br />
C 2<br />
H
182 Nicolas Paparoditis and Olivier Dissard<br />
e plotting<br />
P 1<br />
C 1<br />
Figure 3.3.11<br />
Theoretic matching error (pixels)<br />
0.6<br />
0.4<br />
0.2<br />
–0.4<br />
–0.6<br />
O 1<br />
B<br />
P<br />
e alti<br />
e plani<br />
This bias is due to the interpolation process and to the interpolation<br />
function. We still have a bias if we do not use a sub-pixel estimator.<br />
Indeed, if we do so we still implicitly use an interpolation function which<br />
is the nearest-neighbour interpolator.<br />
With the sub-pixel estimation process, the maximum bias occurs for a<br />
sub-pixel translation of 0.25 pixels. The amplitude of this bias depends<br />
O 2<br />
P 2<br />
C 2<br />
e match<br />
0 0.2 0.4 0.6 0.8 1<br />
–0.2<br />
Parallax fractional part (pixels)<br />
Figure 3.3.12 Intrinsic matching errors.<br />
H<br />
sub-pixel match<br />
pixel match
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
on the interpolation function. It can range from 0.15 to 0.25 pixels. This<br />
does not mean that the function giving the best theoretical precision is the<br />
best in practice. In general, the simpler the interpolation functions (using<br />
the smaller neighbourhoods) the more robust they are for real image digital<br />
matching. From now on, we will consider that the average matching precision<br />
is of a quarter of a pixel.<br />
In the case of digital manual stereo-plotting, the matching precision is<br />
difficult to evaluate statistically because of all the factors to be taken into<br />
account: the radiometric (contrast), geometric (punctual, linear, corner,<br />
etc.) and 3D context characteristics of the feature to be plotted, the image<br />
quality (noise, blur, etc.), the stereoscopic acuity and the tiredness of the<br />
operator, the optical quality of the stereo display, the sub-pixel possibilities<br />
of the stereo display (sub-pixel image displacement, etc.), and the<br />
plotting methodology. But in fact, most of these difficulties are present<br />
with automatic matching techniques too.<br />
3.3.3 Stereo matching from object space (SMO)<br />
Nicolas Paparoditis<br />
Photogrammetrists may express the matching problem for the 3D localization<br />
in a completely different way. In fact, the other way round: ‘given a<br />
(X, Y) in object space what is the corresponding z?′. As we can see in Figure<br />
3.3.13, for a given (X, Y) if we change the z value, the displacements of the<br />
corresponding image points are implicitly constrained along the curves<br />
which are the projection of the planes lying upon each nadir axis and<br />
(X, Y, z). Looking for the best z, is thus looking for the best image match<br />
along these ‘nadir’ curves. Thus as for the image matching process guided<br />
from image space, we build a correlation profile but instead of computing<br />
all correlation scores for all the possible parallaxes along the epipolar line<br />
we compute all correlation scores for all possible z values in a [z min, z max]<br />
search space interval. The best elevation estimate is given by:<br />
where<br />
and<br />
Loc I1 : ℜ 3 → I 1 (x, y, z) → (i 1, j 1)<br />
Loc I2 : ℜ 3 → I 2 (x, y, z) → (i 2 , j 2 )<br />
3D data acquisition from visible images 183<br />
zˆ(X, Y) ArgMax C2V2(Loc (X, Y, z)), V I2 1(Loc (X, Y, z)) ,<br />
I1<br />
z∈[zmin,zmax] (3.14)<br />
are the object to image localization functions. One should remark that<br />
with this stereo method guided from object space, the image matching and
184 Nicolas Paparoditis and Olivier Dissard<br />
Figure 3.3.13<br />
x<br />
z<br />
y<br />
O 1<br />
C 1<br />
the 3D localization processes are performed at the same time. The elementary<br />
sampling dz along the vertical line can be determined to be:<br />
dz , (3.15)<br />
where r0 H/f is the ground pixel size.<br />
H<br />
4B r0 Resampling correlation templates<br />
Due to the perspective deformations and to the relative rotation between<br />
image planes, the image patches corresponding to a flat square ground<br />
patch ds do not overlap. Thus to obtain the best and the most discriminating<br />
correlation score between those patches, the images are locally<br />
resampled.<br />
Indeed, as shown in Figure 3.3.14, we build two ortho-templates centred<br />
on (X, Y). The image resolution of these templates is equivalent to that of<br />
the images. For each pixel of these two ortho-templates we determine the<br />
3D position on the ground patch, and we assign to each of them the greylevel<br />
value for the corresponding image position. If our local flat ground<br />
assumption is correct these should geometrically be exactly overlapping.<br />
Stereo matching from object space is the proper way of obtaining automatically<br />
3D measurements in a stereo display. Indeed, the observation<br />
dz<br />
(X, Y)<br />
O 2<br />
C 2<br />
Z max<br />
Z min
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 3.3.14<br />
x<br />
space in a stereo display is directly the object space. The operator can, in<br />
real time, obtain the elevation corresponding to a given plotter position<br />
(X, Y) or can plot a flat line and have the 3D profile along this line, e.g.<br />
for volume calculations or for road elevation profile estimation.<br />
3.3.4 Difficulties due to image differences<br />
Nicolas Paparoditis<br />
z<br />
P 1 (i 1, j 1, g 1)<br />
y<br />
3D data acquisition from visible images 185<br />
O 1<br />
C 1<br />
(X, Y)<br />
ds(z)<br />
Classical template matching suffers from a certain number of defaults, due<br />
to the non-respect of the assumptions it relies upon, that is to say the<br />
geometric and radiometric similarity of templates centred on homologous<br />
points and the discrimination power of correlation scores computed on<br />
the radiometric templates. Sometimes, the landscape properties and the<br />
geometric configurations of the image views are such that these assumptions<br />
are not respected, and thus the similarity, existence and uniqueness<br />
constraints for homologous points solution are violated:<br />
• On homogeneous areas and periodic structures. The template radiometric<br />
content does not ensure a possible discrimination between all<br />
possible matches that fall in this area. The uniqueness constraint is<br />
violated. Having a higher image quality or a larger window definitely<br />
reduces the ambiguities for homogeneous areas. Periodic structures<br />
are a problem when the spatial periodicity orientation is along the<br />
O 2<br />
C 2<br />
P 2 (i 2, j 2, g 2)<br />
Ortho-template of image 1 1<br />
Ortho-template of image 1 2
186 Nicolas Paparoditis and Olivier Dissard<br />
Figure 3.3.15 Homogeneous surfaces and repetitive structures.<br />
baseline. Having another view not on the same stereo baseline will<br />
considerably reduce the ambiguities. (See Figure 3.3.15.)<br />
• On hidden areas. In urban areas, for instance, all points observable<br />
in an image do not always exist in the other image. Especially in urban<br />
areas some zones are not seen on both images. Template matching is<br />
a low level process. Even if there are no possible matches for a given<br />
(i 1 , j 1 ) or for a given (X, Y), we will still have a maximum peak in the<br />
correlation profile or surface which will lead to a false match. The<br />
existence constraint is violated. A way of bypassing this problem, is<br />
to have a higher stereo overlap (or more generally along or/and across<br />
track multi-stereo views) so that at least every point of the scene can<br />
be seen in at least two images, or to use multi-stereoscopic techniques.<br />
(See Figure 3.3.16.)<br />
• On non-overlapping areas. The incomplete overlapping of the stereopair<br />
is one of the problems that occurs in digital stereo matching.<br />
Indeed, many points of both images having no homologous solutions<br />
will lead to problematic mismatches. The existence constraint is<br />
violated.<br />
Figure 3.3.16 Mobile vehicles and hidden areas.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4111<br />
45111<br />
3D data acquisition from visible images 187<br />
Figure 3.3.17 Differences in size and in shape of two real homologous patterns<br />
on steep sloped surfaces.<br />
• On steep sloped surfaces. The image pattern shapes corresponding to<br />
a sloped planar ground patch in object space are geometrically not<br />
alike. These deformations are accentuated when the slope rises. Thus<br />
square image templates centred on homologous points have less<br />
probability of looking alike when their size increase. The similarity<br />
constraint is violated. A solution is having higher quality images so<br />
that we could take smaller windows, or coping with the deformations<br />
by using adaptive shape windows. (See Figure 3.3.17.)<br />
• Areas around 3D discontinuities. When a template overlaps a 3D elevation,<br />
as all the pixels contribute to the matching process, the elevation<br />
determined in these areas will be a ‘mean’ of all 3D elevations inside<br />
the templates weighted by the contrast distribution. This will lead to<br />
a smoothing of the 3D relief morphology or to a spatial shift of the<br />
3D structure of one half of the window size, as shown in Figure 3.3.18.<br />
We can overcome this problem using small window sizes thus assuming<br />
again high image quality, or using adaptive shape windows.<br />
• Non-Lambertian surfaces. This happens for areas that have anisotropic<br />
surface reflection properties. If we observe this surface from different<br />
viewpoints, the radiometric measurements can be radically different.<br />
When the images are not saturated, and if the non-Lambertian effects<br />
are due relatively to a smooth surface compared to the ground pixel<br />
size, template matching can still be used if the texture vectors are<br />
centred to correct the radiometric bias between both homologous<br />
templates. When some areas are saturated (no details) in one of<br />
the images or when the non-Lambertian effects are due to 3D microstructures,<br />
for instance a cornfield seen under two different viewing<br />
angles, there is no satisfactory way of finding a good solution other
188 Nicolas Paparoditis and Olivier Dissard<br />
(i, j) (i′, j′)<br />
(i′′, j′′)<br />
Figure 3.3.18 Matching areas around 3D changes. If we are looking for the<br />
homologous template of (i, j) along the epipolar line, the best<br />
correlation score will be obtained for (i′′, j′′) instead of (i′, j′). Thus<br />
the parallax/elevation associated to (i, j) will be the one of the<br />
rooftop.<br />
than rejecting the matches for these areas, unless we consider another<br />
stereopair if we have a survey with a high stereo overlap, or if we are<br />
in any multi-stereoscopic configuration.<br />
• Moving vehicles and shadows. Matching problems can occur due to<br />
the fact that stereopairs are not instantaneous. Indeed, features like<br />
vehicles and shadows can move from one image to the other. Some<br />
vehicles appear or disappear, thus the existence constraint is violated.<br />
For some vehicles moving along the baseline, the matching process<br />
will generate aberrant corresponding 3D structures integrating the 3D<br />
and the movement parallax. For satellite or aerial across track stereopairs,<br />
shadows have moved thus the matching process can reconstruct<br />
3D shadow limits at non-realistic elevations.<br />
3.3.5 Coping with the problems: analysis of the correlation<br />
profile<br />
Nicolas Paparoditis<br />
One of the easiest solutions to coping with these problems is to be able<br />
to detect and reject all the matches where we think that the probability<br />
of encountering a problem is high. We can study the morphologic structure<br />
of the correlation profile and verify that it respects the properties of<br />
a good match, i.e. high score for the maximum peak, large correlation<br />
principal peak, no high secondary peaks. Many correlation profiles whatever<br />
the area do not respect this model. Indeed, some good matches have<br />
low and some others high correlation scores, and some false matches have<br />
high correlation scores. This is thus not an acceptable solution if we want
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
this process to be universal. Of course, this kind of process can be applied<br />
as a last resort when no particular solutions can be found for each one<br />
of these area problems.<br />
3.3.6 Coping locally with the problems: adaptive window<br />
shapes and sizes<br />
Nicolas Paparoditis<br />
3D data acquisition from visible images 189<br />
The major problem of correlation template/area-based matching techniques<br />
is due to their shape rigidity (square or rectangular) and their fixed sizes.<br />
On the one hand, small templates enable an improved location of the good<br />
matches. On the other hand, larger windows allow a better discrimination<br />
between all possible matches. The problem is that we are confronted<br />
with a classical detection/localization compromise induced by the rigidity<br />
of the templates. Templates’ sizes and shapes should thus change automatically<br />
and adapt themselves according to the changes in the landscape.<br />
We will now explain how this can be done.<br />
Coping with homogeneous areas and periodic structures<br />
For the SMI method and the homogeneous areas, the window size can be<br />
adapted to the local variance of the slave image contents having some<br />
knowledge about the image noise distribution. When areas are homogeneous,<br />
windows will be large, and when the areas are rich in texture,<br />
windows will be small. For periodic structures, we can choose the size of<br />
the template so that the auto-correlation of the texture in its own neighbourhood<br />
along the base line is mono-modal. This would mean that the<br />
texture comprised in the template is no longer a periodic pattern.<br />
These adaptive size techniques are not possible with the SMO method.<br />
Indeed, we would correlate templates of different sizes and this would lead<br />
to heterogeneous correlation scores for all the elevations, thus choosing<br />
the correlation peak could lead to a false match. And it is difficult weighting<br />
the correlation scores according to template sizes. Nevertheless, one can<br />
verify that the templates corresponding to the correlation peak are textured<br />
enough to start, if necessary, another matching process on larger templates.<br />
Another solution is to take a large window of fixed size for all the image<br />
and to weigh the grey-level values by a Gaussian distribution centered on<br />
(i1, j1). The interest of this technique is giving more importance to the<br />
pixels closer to (i1, j1). Nevertheless, if the area is globally homogeneous,<br />
an image feature inside the template window, even if it is relatively far<br />
away from (i1 , j1 ), will allow a possible discrimination in the matching<br />
process. Besides, the estimated disparity will be the one corresponding to<br />
the feature and not to the central point of the template window.
190 Nicolas Paparoditis and Olivier Dissard<br />
Coping with hidden parts<br />
For the SMI method, using the reciprocity constraint is an easy and efficient<br />
way of filtering false matches. We verify that the match of (i 2 , j 2 ),<br />
homologous of (i 1, j 1), is also (i 1, j 1). This process is often called crossvalidation.<br />
This filtering is very discriminating for false matches due to<br />
hidden areas. Indeed, if (i 1, j 1) belongs to a hidden area, it is most probable<br />
that the homologous of (i 2 , j 2 ), estimated as the homologous of (i 1 , j 1 ), will<br />
not be (i 1 , j 1 ).<br />
For the SMO method, we can do a similar process. For the elevation<br />
and the (i 1, j 1) and (i 2, j 2) corresponding to the correlation peak, we can<br />
verify with the SMI method that (i 2, j 2) is the homologous of (i 1, j 1) and<br />
vice versa.<br />
Coping with discontinuities<br />
We can adapt the templates locally if we have a priori information of<br />
where those 3D structures occur. If we do not have any maps, the only<br />
way of obtaining this information is by computing contour maps derived<br />
from the images assuming that a 3D discontinuity is often characterized<br />
in the images by a radiometric contrast. We can locally adapt the template<br />
shapes to the contours so they do not integrate patterns on the other side<br />
of the contours which could be areas at different elevation (Paparoditis<br />
et al., 1998). Cutting the template to produce a mask will of course reduce<br />
the textural content of the template, thus this process can only be carried<br />
out with large enough templates. These kinds of processes are half-way<br />
between area-based and feature-based matching techniques.<br />
With the SMI method, for (i 1, j 1) we limit the template to all the pixels<br />
that are connected to (i 1, j 1), i.e. there is a non-contour crossing path<br />
joining, as shown in Figure 3.3.19(a). If a contour crosses continuously<br />
the template there will be no possible path joining the pixels on the other<br />
side of the contour. If a part of a structure corresponding to a 3D discontinuity<br />
is missing in the contour detection, the adaptive template will be<br />
the classical template itself. A solution is to constrain the path to be straight<br />
lines, as shown in Figure 3.3.19(b). The counterpart of this solution is<br />
that small open contours inside the template will mask all the pixels behind<br />
them and limit the number of pixels intervening in the calculation of the<br />
correlation score. A better solution to take into account those open contours<br />
is to weigh, as shown in Figure 3.3.19(c), all the grey-levels pixels given<br />
by the Figure 3.3.19(a) process by the inverse of the geodetic distance<br />
(Cord et al., 1998) which is given by the length of the shortest path joining<br />
a pixel inside the template and (i 1, j 1).<br />
If (i 1, j 1) is itself on a contour, what should the template be: the right<br />
side or the left side of the contour? Neither in fact, the best solution in<br />
this case is feature-based matching, that is to say matching the contour
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
(a) (b) (c)<br />
Figure 3.3.19 In the centre the pixel (i1 , j1 ), in black contours pixels, in grey the pixels belonging to the<br />
adaptive template, in white the pixels which are not taken into account in the correlation<br />
score. (a) template mask with path existence; (b) template mask with straight line paths;<br />
(c) template mask with path existence weighted by the inverse of the geodetic distance.
192 Nicolas Paparoditis and Olivier Dissard<br />
itself. One solution is to look for all the contours in image I 2 along the<br />
epipolar line. Nevertheless, we need to remove ambiguity from all possible<br />
matches by considering the intrinsic information of the contour (we can<br />
verify that the contour orientations are the same) and the geometric and/or<br />
radiometric context around the contour (many solutions are possible). To<br />
add some robustness to this process, we can average, or take the median<br />
value of all disparities for all pixel edges locally along the contour line on<br />
each side of (i 1, j 1) assuming that the disparity is locally stable. These<br />
feature-based matching techniques are often not valid when the contours<br />
are not stable from one image to the other. This is the case for instance<br />
for building edges in urban areas, which are most often not characterized<br />
in the same way in both images.<br />
With the SMO method, we consider the intersection of both image adaptive<br />
template shapes for a given elevation. The interest of this, compared<br />
to that of the SMI method, is that if a part of a 3D contour is missing in<br />
one of the two images, as long as the contours are complementary in both<br />
images this adaptive process will be efficient. Meanwhile this process is to<br />
be carried out carefully, as the number of pixels intervening in the correlation<br />
score calculation change all along the correlation profile.<br />
Coping with geometrical distortions due to slopes<br />
When image quality is good, small templates can be chosen to limit the<br />
distortion effects of sloped relief. In this case we may transform geometrically<br />
the homologous template (when guided from image space) or both<br />
templates (when guided from object space). We can model locally the relief<br />
by a polynomial function of a given order. Then the matching problem<br />
can be seen as finding the optimal set of coefficients giving the highest<br />
correlation score. When the order of the function increases, it is very difficult<br />
finding a robust solution and the process is much longer in time. So<br />
it is preferable to make stricter assumptions on the surface, i.e. supposing<br />
that the surface is locally planar.<br />
Under these assumptions, for the SMI method we can determine for the<br />
square template around (i1 , j1 ) in the reference image, the grey levels of<br />
the homologous template (same size and shape) by intersecting for each<br />
one of its pixels the corresponding 3D ray with a given patch hypothesis<br />
to determine the 3D corresponding point in object space. Given this 3D<br />
point we can determine the corresponding point in I2 and interpolate the<br />
corresponding grey-level value we were looking for and then determine<br />
the corresponding image position and grey level (by interpolation) in the<br />
other image. For all the set of ground patches hypotheses, we can construct<br />
all the homologous templates and thus all the similarity scores. (See Figure<br />
3.3.20, colour section.)<br />
When guided from object space, we can calculate the correlation scores<br />
for all the ortho-templates built for each possible ground patch (z, , )
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 3.3.20 (a) Left image of an 1 m ground pixel size satellite across track<br />
stereopair; (b) right image; (c) DSM result with a classical crosscorrelation<br />
technique; (d) DSM using adaptive shape windows.
194 Nicolas Paparoditis and Olivier Dissard<br />
Figure 3.3.21<br />
x<br />
centred on (X, Y) as shown in Figure 3.3.21, (resp. ) is the angle<br />
between the surface normal and x (resp. y). To limit the search-space<br />
possibilities we can make some assumptions on the maximum slope, e.g.<br />
the slope cannot reasonably exceed 45°. One should remark that implicitly<br />
the 3D entity we are looking for is not the 3D point on (X, Y) but an<br />
entire surface patch around (X, Y).<br />
All these adaptive matching techniques can be mixed together. For<br />
instance, the slope and the 3D discontinuity techniques are to be mixed<br />
together if one wants to determine a 3D measurement of a point on a<br />
sloped roof next to the roof’s edge. The combination of these two techniques<br />
is a strategy guided by the type of landscape to be mapped.<br />
3.3.7 Strategies for a stereopair DSM generation<br />
Nicolas Paparoditis<br />
z<br />
y<br />
O 1 O2<br />
C 1<br />
(X, Y)<br />
We discuss here the strategies of integration of different methods and techniques<br />
to build a regular grid DSM describing the observable surface. We<br />
do not propose an ideal and universal strategy of techniques integration<br />
but different ways of integrating the techniques and the problems related<br />
to their integration. Indeed, the strategy will depend on the type of land-<br />
→<br />
n<br />
→<br />
ds(z, n(α, β))<br />
C 2
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 195<br />
scape from natural smooth terrain to complex and dense urban areas, on<br />
the different image resolutions from low to very high, on the data acquisition<br />
conditions, and on image quality, etc. There is possibly a universal<br />
strategy, but it has not yet been found. Even if this strategy exists, some<br />
simpler strategies will be lighter in terms of computing times and efficient<br />
enough for a given set of data on a given landscape.<br />
Basic matching entities and spatial sampling<br />
Can we derive a DEM from a sparse set of 3D points corresponding to<br />
some matched image features entities, i.e. interest points, contour points,<br />
etc. in order to reduce the computing times? The underlying assumption<br />
would be that any relief point could be determined from the nearest 3D<br />
computed features by an interpolation process. Image features are sparse<br />
and do not sample the relief well enough to describe it correctly. Moreover,<br />
these entities are not stable and discriminating enough (especially in the<br />
case of urban scenes) thus leading to ill matches. The denser the entities,<br />
the finer the surface will be sampled. Points or surface patches are thus<br />
the basic matching entities for the DEM matching process even though<br />
image features can help in the process as we will see later on.<br />
In the case of smooth natural relief, surface patches entities seem to be<br />
the most relevant to describe the surface. If we assume the relief to be<br />
smooth enough and planar locally with respect to the size of the patches,<br />
the DEM spatial sampling need not be that of ground pixel size. The DSM<br />
sampling should be constrained by our knowledge of the surface regularity.<br />
In consequence if the surface is smooth the ground size of the surface<br />
patches could be equivalent to the spatial sampling itself. Nevertheless, if<br />
we want the DSM to describe the surface in the most precise way, not<br />
only should the elevation be stored for every DSM grid point. Indeed, for<br />
many applications, the slope information is more or as relevant than the<br />
elevation one. So, the surface patch normal vector could also be stored<br />
for every grid point to improve the precision of both elevation and slope<br />
interpolation processes for a point inside the grid.<br />
For ‘rougher’ surfaces, as in urban areas, a spaced grid of rectangular<br />
shape surface patches would not be adequate as the patches should be<br />
very small to describe the surface correctly. Thus the texture contents of<br />
the corresponding image templates will often not be rich enough to be<br />
entity discriminating. For urban surfaces, we prefer using a denser sampling,<br />
i.e. the process can be carried out for each pixel of image I1 with the SMI<br />
method or for every grid point (X, Y) of the DSM with the SMO method.<br />
In the latter method, the DSM grid spatial sampling should be that of the<br />
size of the smallest objects we would like to represent. Meanwhile, this<br />
does not mean that the spatial resolution of the DSM is equivalent to the<br />
ground pixel size as the measured elevations are a ‘mean’ of all real elevations<br />
inside the template.
196 Nicolas Paparoditis and Olivier Dissard<br />
Initializing the matching process<br />
A gross approximation or knowledge of the surface helps to reduce considerably<br />
the range of the search space intervals to diminish the computing<br />
times and to reduce the matching ambiguities probability. The initialization<br />
strategies are various.<br />
Hierarchical matching<br />
Hierarchical matching processes (Grimson, 1983) first aim at matching a<br />
very restricted subset of image entities i.e. characteristic image features such<br />
as, for example, interest points or contours and use these matches to<br />
initialize the matching of all the other image entities. The matching hypotheses<br />
for all the points lying between the matched features are searched for<br />
in a disparity/elevation search space interval framed by those of the matched<br />
features. This process is efficient when the feature entities are stable.<br />
Using the tie points<br />
A gross surface can be constructed using the tie points that have been used<br />
to determine the relative orientation of the images in the sensor viewing<br />
parameters determination process. The 3D samples corresponding to the<br />
matched tie points can be triangulated to construct a TIN surface that can<br />
be used to initialize the DSM generation. The image can be extremely<br />
dense and thus the gross surface can be quite close to the real surface.<br />
Multi-resolution matching<br />
Multi-resolution matching can be seen as a form of hierarchical matching.<br />
We first build for each image a multi-resolution pyramid as shown on<br />
Figure 2.5.3 using for instance a smoothing kernel filter, e.g. a Gaussian<br />
kernel or a wavelet decomposition. The matching starts from the lower<br />
level of the pyramid, (lowest resolution image) where the disparity vectors<br />
and the range of the search space are much smaller. The surface elaborated<br />
at this resolution is used to initialize the matching at the next step<br />
of the pyramid and so on. The major advantage of this strategy is to lower<br />
in a consequent way the computing times. This matching strategy is valid<br />
when the solution (DSM) provided for a given resolution is close to the<br />
surface observable at the highest resolution. It is thus valid for smooth<br />
surfaces but not for surfaces encountered in urban areas.<br />
Object space vs. image space method<br />
When the matching process is guided from image space, the 3D points<br />
corresponding to the matches of points located on a regular grid in slave
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 197<br />
image space have an irregular distribution in object space. Thus the generation<br />
of a spatial regular grid DEM goes through the blind resampling/<br />
interpolation of the irregularly distributed set of 3D samples, which<br />
inevitably leads to artefacts and alters the quality of elevations computed<br />
for all the DEM grid points. It is thus highly preferable determining the<br />
elevations of all the DEM grid points with the SMO method.<br />
As a counterpart, some drawbacks in terms of practical implementation<br />
are to be mentioned. If we generate the DEM by computing the elevations<br />
sequentially for all the points some basic operations are highly redundant,<br />
thus the computing times will unnecessarily increase. Indeed, we mentioned<br />
earlier that the correlation score was applied for a given (X, Y) and for a<br />
given z to the ortho-templates generated by the local resampling of the<br />
images. Neighbouring points in the grid will have for the same z overlapping<br />
ortho-templates. Thus the resampling process for all pixels inside<br />
the overlap will be carried out in a redundant way. And the larger the<br />
windows the higher the redundancy. How can we avoid this? The solution<br />
is to resample both images in object space once for all for a given z.<br />
In other words we build two ortho-images (which have the same geometry<br />
as our DEM) assuming that surface is flat of elevation z. If the relief<br />
around a given (X, Y) is at an elevation z the ortho-templates extracted<br />
from these ortho-images are centred and (X, Y) should be alike. Thus for<br />
this given z and for all the (X, Y) of the DEM grid we can compute all<br />
the correlation scores.<br />
In practice, the determination of the elevations for each DEM grid point<br />
can be done in two ways. The first one consists in initializing the DEM<br />
elevations to the lowest possible z and determining all the correlation scores<br />
for this z and for every point of the DEM grid with our global resampling<br />
and correlation process (we store the correlation scores in a corresponding<br />
buffer grid). For the next possible z we globally process the correlation<br />
scores in the same way. All DEM grid points where the new correlation<br />
scores are higher than those computed with the previous z have their<br />
elevation and correlation scores updated. We continue this process until<br />
we reach the highest possible z. If the storage capacities are limited, we<br />
can cut the global DEM into smaller ones and treat each of them independently<br />
and sequentially. The first interest is that we can verify that the<br />
correlation profile is a good-looking one. The second is that the correlation<br />
profile peak does not always identify the real solution which appears<br />
in the correlation profile as a secondary peak. Keeping all the information<br />
leaves us the possibility of choosing the best solution inside the peak set<br />
with a higher-level processing. This is what we are going to develop now.<br />
Improving the process by looking for a global solution<br />
We have up to now addressed the problem of independent 3D entity measurements.<br />
In the case of dense DEM measurements, we have generated a
198 Nicolas Paparoditis and Olivier Dissard<br />
dense cloud of independent 3D entities that are supposed to describe a<br />
surface. These points should describe a ‘possible’ surface and neighbouring<br />
measurements should be coherent to a certain extent. Indeed, a 3D<br />
measurement cannot be completely decorrelated from its neighbouring<br />
measurements. Even when discontinuities occur their should be a<br />
correlation between all neighbours on each side of the break. We can thus<br />
verify locally or more globally the internal coherence of all the measurements<br />
assuming some modelled hypotheses or constraints on the surface<br />
basic properties.<br />
Detecting erroneous 3D samples<br />
A simple technique often applied to detect erroneous samples is based on<br />
the ordering constraint also called visibility constraint. This constraint,<br />
based on the assumption that the observable surface is a variety in the<br />
form of a single-valued function z f(x, y), verifies (for the SMI method)<br />
that three image points in a given order on an epipolar line also follow<br />
the same ordering on the conjugate epipolar line.<br />
This constraint can also be applied, in object space, to the surface<br />
samples. Indeed, every 3D sample has been determined by matching homologous<br />
points thus supposing implicitly that these points are visible in both<br />
images. We can thus verify, with a z-buffering algorithm, that every 3D<br />
sample is directly visible in both images and not hidden by any other<br />
3D samples. If some samples are hidden we have some evidence of the<br />
existence of erroneous samples. As shown in Figure 3.3.22 there is an<br />
obvious incompatibility between P1 and P2. P 1 or P 2? O 1 O 2 P 2<br />
C 1<br />
P 1<br />
Figure 3.3.22 Mapped surface incoherence detection.<br />
P 2<br />
C 2<br />
H
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 199<br />
In the case of natural relief, we can verify that the surface is locally,<br />
smooth, planar and continuous. For a 3D patch process, we can verify<br />
that neighbouring 3D patches are connex. If one of the patches is not at<br />
all correlated with its neighbours, the patch is discarded and can be replaced<br />
by a patch fitting with the neighbouring ones. For a 3D point process, we<br />
can determine a plane fitting locally the 3D neighbour samples by a least<br />
squares estimation technique and verify that the considered point is close<br />
enough to this plane. We can still manage this process if a discontinuity<br />
occurs assuming that this discontinuity is characterized by a contour in<br />
one of the images. Indeed, we can transpose the 3D samples in image<br />
space and we can consider only the neighbouring samples that are on the<br />
same side of a contour. Besides, if a local connected set of 3D entities is<br />
erroneous, i.e. the errors are structured and spread, this process rejoins<br />
the ill-posed problems.<br />
Finding the best 3D surface in a global way<br />
A ‘clean’ solution to find the most ‘probable’ surface is keeping for all the<br />
entities all or a subset of the best (the highest correlation peaks) matching<br />
hypotheses and finding the best path or surface fitting the ‘fuzzy’ measurements<br />
by an optimization technique. We can evaluate the realism and the<br />
‘probability’ of each possible path and retain the best one given a cost<br />
function integrating constraints on the surface behaviour. Figure 3.3.23<br />
z<br />
C 1<br />
O 1<br />
O 2<br />
Structured erroneous hypotheses<br />
C2 1st best match peak<br />
2nd best match peak<br />
3rd best match peak<br />
x<br />
best ‘fuzzy’ path<br />
path given by the best<br />
matching hypotheses<br />
Figure 3.3.23 A virtual example of possible path optimization of the best 3D<br />
samples hypotheses given by the SMO method. In black we have<br />
the path given by the independent best matching hypotheses.<br />
In dotted line, we have the best ‘fuzzy’ path assuming some<br />
constraints, i.e. the surface is continuous and smooth with no steep<br />
slope.
200 Nicolas Paparoditis and Olivier Dissard<br />
shows an example of a global optimization with a continuity constraint<br />
where we have kept for each (X, Y) along a 3D ground profile, the three<br />
most probable z values corresponding to the three highest correlation peaks<br />
obtained with the SMO method.<br />
The optimization techniques that can be applied are numerous, i.e. relaxation,<br />
least squares, dynamic programming, etc. We will not describe these<br />
techniques here, one can find them in any good signal or image processing<br />
book. What differentiates these techniques is the domain on which the<br />
optimization is carried out. Relaxation is usually applied to small neighbourhoods<br />
around the samples. Dynamic programming is applied to<br />
samples along linear ground (as shown in the example of Figure 3.3.23)<br />
or image (e.g. epipolar lines) profiles. The larger is the domain, the higher<br />
are the chances of finding a ‘correct’ global solution but the higher is the<br />
optimization complexity and the computing times. Besides, if we choose<br />
local domains, we have to ensure that all the connected domains are<br />
coherent thus leading us to another optimization problem.<br />
The optimization processes can be carried out in the same way in the<br />
object space, i.e. on the 3D samples hypotheses or in image space on the<br />
matching hypotheses. One of the major interests of these processes besides<br />
improving globally the results is that we can use windows of smaller sizes<br />
and have a better local and morphological rendering of small 3D structures.<br />
Moreover, using the windows that are as small as possible is a<br />
guarantee that the neighbouring measurements are the ‘more’ independent<br />
as possible thus lowering the probability of having structured erroneous<br />
areas that could mislead the optimization process.<br />
When the surfaces are not continuous but piecewise continuous as in<br />
urban areas, the optimization process should take into account the location<br />
of the 3D break lines to authorize locally a depth or a disparity<br />
discontinuity. As we do not have any a priori information on their 3D<br />
location, the only information we can inject in the process is the location<br />
of image contours, which are the projections of possible 3D breaks (March,<br />
1989). We can then apply a continuity constraint on the hypotheses<br />
(brought back in image space) on each separate side of the contour. As<br />
we said earlier, in urban areas, contours corresponding to building limits<br />
do not appear in all the images at the same time, but most of the time a<br />
given building edge appears contrasted in at least one of the images. Thus<br />
if we want the break process to be efficient we have to consider the optimization<br />
on all the contours of all the images at the same time. Nevertheless,<br />
one should have in mind that this process can in some cases generate<br />
virtual depth discontinuities around non-3D contours.<br />
Figure 3.3.24 An example of the management of hidden parts(a): in (b) the left<br />
image; (c) the raw DSM, with (in black) the areas where the correlation process<br />
failed. Considering that these zones are generally on the ground, by subtracting<br />
the DTM (e), obtained by smoothing the data acquired between the raised<br />
structures, one gets a DEM (d) with only minor imperfections.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
O 1<br />
C 1<br />
C 2<br />
O 2<br />
(a)
202 Nicolas Paparoditis and Olivier Dissard<br />
It is necessary to discard, before the optimization process, all the grid<br />
points that have no significant correlation profile. Indeed, these points<br />
could correspond to areas that are hidden due to occluding structures. Thus<br />
looking for a path in these areas would have no sense and would lead in<br />
all cases to an erroneous solution. Should these points be a posteriori<br />
blindly interpolated in the DEM generation process? Certainly not! This<br />
would lead us to odd-looking surfaces. The hole-filling strategy should be<br />
decided and guided by the operator. The alternative solution if we desire<br />
a fully automatic process would be to have the existence of a solution<br />
inside the matching hypotheses for every landscape point. This means that<br />
every landscape point should be seen in at least two images of the survey.<br />
In the case of urban surveys, this implies a high stereo overlap (along and/or<br />
across-track). The optimization process should then be carried out on all<br />
the 3D sample hypotheses accumulated from all the possible stereopairs<br />
(Roux et al., 1998). Another alternative is to consider that missing<br />
areas are on the ground. Identifying and removing all the raised structures<br />
through a joint image/analysis will leave inside the DSM only the ground<br />
elevation samples. Thus the elevation values for the initial missing areas can<br />
be interpolated from all the set of ground elevation samples. (See Figure<br />
3.3.24, colour section.)<br />
3.3.8 An example of strategy for complex urban landscapes:<br />
dynamic programming applied to the global and<br />
hierarchical matching of conjugate epipolar lines<br />
Olivier Dissard<br />
We will describe here a global matching process based on dynamic<br />
programming that integrates contours to model the 3D discontinuities.<br />
This scheme (Baillard and Maître, 1999; Baillard and Dissard, 2000) has<br />
given very good results on urban, suburban, and rural landscapes.<br />
What is dynamic programming in the case of epipolar images<br />
matching?<br />
For the purpose of stereo image matching, dynamic programming should<br />
be considered as the search for the ‘best’ path between two known pairs<br />
of homologous points, to derive the 3D profile between these two points.<br />
‘The best’ is considered in relation to an energy function that assesses the<br />
resemblance between each of all the possible pairs of points.<br />
Considering the SMI point of view, epipolar resampled images represent<br />
an interesting workspace: considering each pair of homologous lines,<br />
the aim becomes to search for a ‘best’ path between the two origins and<br />
ends (or two known pairs of homologous points) in the space defined by
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 203<br />
the right and left epipolar lines and filled with a similarity function (see<br />
Figure 3.3.25). All the explanations that follow are based on a ‘pixel-topixel’<br />
but not a sub-pixel matching.<br />
Similarity function<br />
A necessary condition for the similarity function is that the computation<br />
must be comparable everywhere: for example, adaptive windows for the<br />
computation of cross-correlation coefficient do not provide comparable<br />
values because the height of the correlation peak depends on the size of<br />
the window. Some computation tricks reduce the influence of window size<br />
without solving it. Thus, in these cases there can not be an objective optimized<br />
path.<br />
The similarity assessment used for the matching in Figures 3.3.25 and<br />
3.3.26 is nothing but the simple cross-correlation coefficient on a neighbourhood<br />
with constant size. It does not assess the resemblance of the<br />
local radiometry, but the resemblance of the radiometry change inside the<br />
neighbourhood. We will see later how to take radiometric resemblance<br />
into matching with dynamic programming.<br />
Possibilities for paths<br />
On the assumption that objects on both epipolar lines should be found in<br />
the same order (which is quite the reality, except mobile objects or perspective<br />
effects on lateral sides of the epipolar pair between pylons or trees<br />
and ground texture), each possible path that goes through a point comes<br />
from only three directions (Figure 3.3.25). It means in fact that there is<br />
no possibility of going back on one image when going forward on the<br />
m–1<br />
m<br />
Right epipolar line<br />
n–1<br />
2<br />
3<br />
n<br />
1<br />
Left epipolar line<br />
Figure 3.3.25 Possible paths for matching with epipolar lines.
204 Nicolas Paparoditis and Olivier Dissard<br />
other one. This simple but quite robust assumption makes the computation<br />
possible, for the best path through a point derives from three best<br />
paths with coordinates less than or equal to the considered ones.<br />
The selected path on the point P(n, m) is then:<br />
P(n, m) Max(P(n, m1) Energy(1), P(n1, m1) <br />
Energy(2), P(n1, m) Energy(3)). (3.16)<br />
Let us examine each path:<br />
• Path 2 means that the disparity of (n, m) is the same as the disparity<br />
of (n1, m1): we are on a horizontal 3D surface.<br />
• Path 1 means that we have moved one pixel forward on the right<br />
image, keeping the same position on the left one, disparity has<br />
progressed. If (n, m) belongs to the right path, it means that we are<br />
on a part of the right image that does not appear on the left one:<br />
occlusion, mobile vehicle, specularity, etc.<br />
Path 1 is also for sloped areas: for example, a 1 ⁄2 sloped profile will<br />
alternate Path 1 and Path 2 through the optimized path.<br />
• Path 3 is the opposite of Path 2, regarding left and right lines.<br />
Into the algorithm, point (n, m) will be assigned with the value of P(n, m)<br />
and O(n, m) which is the provenance (1, 2 or 3) of P(n, m):<br />
O(n, m) Arg(energy) (Max(P(n, m1) Energy(1),<br />
P(n1, m1) Energy(2), P(n1, m) Energy(3))).<br />
(3.17)<br />
Thus P is used for progressing from the beginning point to the end point,<br />
while O allows going back to the beginning point, once the end point is<br />
reached.<br />
Energy function<br />
Let us consider Path 2: it means that disparity is the same between (n, m)<br />
and (n1, m1), in other words, it means that the 3D profile is continuous,<br />
the homology between points n and m is acceptable, therefore the<br />
similarity function must provide an acceptable value.<br />
Considered from an opposite point of view, we will assume that Path 2 is<br />
the right one if the similarity function (the correlation coefficient cc(n, m))<br />
is acceptable, higher than a given threshold ccmin. Path 2 cc(n, m). (3.18)
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 205<br />
What happens when cc(n, m) Path 1 Path 3.<br />
cc min<br />
Thus Path 1 Path 3 (3.19)<br />
2 .<br />
Application<br />
Figure 3.3.27 shows two epipolar lines, a correlation coefficient matrix,<br />
and the optimized path between beginning and end points. This algorithm<br />
is very convenient for urban context where one usually has to deal with<br />
continuous 3D areas, mixed with disparity gaps due to vertical façades.<br />
Mobile vehicles, or opposite contrasts due to specular reflection are overcome<br />
with series of ‘Path 1 Path 3’. However, the algorithm is robust,<br />
for it finds the right path very quickly after such accidents.<br />
Finally the optimized path is ‘symmetric’: the paths are the same if we<br />
go from ‘end’ point to ‘beginning’ point.<br />
There is no necessity in having homologous pairs of points as beginning<br />
and end points, because the incidence is in fact very limited; the right path<br />
is always directly found after the right number of Paths 1 or Paths 3.<br />
Intervals of search can be limited to reduce computation time, for<br />
example with a map of 3D contours: the beginning point is a 3D contour<br />
C1(n1, m1) and the end point is the next one C2(n2, m2): the optimized path<br />
has a length L < n2 n1 m2 m1. Forbidden paths: how to introduce new constraints<br />
For different reasons, some pairs of points cannot be matched. It is<br />
expressed on the matrix by parts or points, through which every Path 2<br />
has a very low energy: instead of computing the correlation coefficient on<br />
these points, we decide that cc
Forbidden area:<br />
Heights > DTM<br />
Right epipolar line<br />
Figure 3.3.26 Restriction of the space search.<br />
Left epipolar line<br />
Forbidden area :<br />
DTM + maximum<br />
height above ground<br />
Figure 3.3.27 Correlation matrix corresponding to the matching of the two<br />
epipolar lines appearing in yellow in the image extracts.<br />
Every grey level with coordinates. (i, j) in this matrix corresponds to the correlation score<br />
and the likelihood of the matching hypothesis between pixel of row number i along the<br />
epipolar line in image 1 and pixel of row number j along the epipolar line in image 2. The<br />
best path found with the dynamic programming within the correlation matrix is superimposed<br />
in yellow. The red lines show some examples of corresponding points found along the path.<br />
The green lines and areas show, respectively, some examples of 3D occultation contours and<br />
the corresponding occluded area in the other image.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 207<br />
radiometric modelling. The examples in Figures 3.3.28 (colour section),<br />
3.3.29 and 3.3.30 are obtained with such a two-step algorithm.<br />
Some parts are forbidden even for occlusion, for example in the case<br />
where we know lower and higher possible altitudes (with a coarse matching,<br />
or by the knowledge of a DTM and a maximum height of the aboveground<br />
objects – trees or buildings). Restriction of the search space greatly<br />
speeds up the process (see Figure 3.3.26).<br />
What matching difficulties does dynamic programming solve<br />
and not solve?<br />
This algorithm is conceived to solve the matching difficulties due to aboveground<br />
vertical façades, and in fact, it solves the problem of mobile vehicles,<br />
when no matching ambiguity appears. Repeated textures, such as parking,<br />
are also solved by dynamic programming, considered as global optimization.<br />
To the contrary, this algorithm suffers from the similarity measure that<br />
takes contrast into account, and thus the borders of 3D objects are unlocalized<br />
with an error equal to the radius of the measure neighbourhood.<br />
There are still matching ambiguities: a mobile vehicle matched with a<br />
parked one – see Figure 3.3.30, confusion with objects on buildings’ walls<br />
occluded on one image, etc.<br />
One can find more details in Baillard and Dissard (2000). Also, see<br />
Figures 3.3.27 and 3.3.28 (colour section).<br />
Figure 3.3.28 Results of the global matching strategy on a complete stereopair<br />
overlap in a dense urban area. The yellow line superimposed on<br />
the two images corresponds to the line we have matched previously<br />
in Figure 3.3.27.
208 Nicolas Paparoditis and Olivier Dissard<br />
3.3.9 Improving the DEM generation in urban areas with<br />
higher quality data: multiple views and digital images<br />
Nicolas Paparoditis<br />
At these levels of resolutions and in a context of dense urban tissue, there<br />
are a considerable number of hidden parts to the extent that a dense and<br />
thorough description of the scene by a stereo analysis is limited (Figure<br />
3.3.29). Besides in urban areas, the classical stereo processing admits some<br />
flaws due to all the matching ambiguities encountered, e.g. mobile vehicles<br />
(Figure 3.3.30), repetitive structures and also due to the poor image<br />
quality. Poor image quality for instance in scanned images is due to noise<br />
added by the scanning process, to the dust, the scratches, and other dirt<br />
present on the silver halide photographs, but also due to the non-linearity<br />
of the silver halide sensibility. Indeed, the signal to noise ratio is dreadful<br />
in dark areas such as shadows, which will thus have poor chances of being<br />
matched if they are hard.<br />
<strong>Digital</strong> images<br />
The quality of the images, i.e. the signal to noise ratio, is the key factor<br />
for the quality of the products that can be derived from these images.<br />
Figure 3.3.29 Missing information in the DSM due to the hidden areas in the<br />
images especially areas around building ground footprints.<br />
Figure 3.3.30 Artefacts due to mobile vehicles. Indeed the stereo-process supposes<br />
that the observed surfaces are still between the two acquisitions.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 209<br />
Indeed, with a digital imaging system with a much higher signal to noise<br />
ratio, the image content within ‘homogeneous’ areas and shadowed areas<br />
is still quite rich, nearly as much as the rest of the image (with just a lower<br />
signal to noise ratio), and will still be fully exploitable. This is a completely<br />
different situation from what happens in classical silver halide pictures,<br />
even if the digitization has been performed with the best possible performances.<br />
(See Figures 3.3.31 and 3.3.32.)<br />
Figure 3.3.31 Impact on homogeneous areas of a higher signal to noise ratio.<br />
For the same building and pixel size: (a) digitized image with<br />
(b) image processing. In (c) CCD digital image (IGN-F camera),<br />
with (d) image processing.<br />
Figure 3.3.32 One can see the poor image content (a) within the shadows due to<br />
the poor sensibility of the film within the dark areas in comparison<br />
with the rich shadow contents (b) inside the digital image thanks<br />
to the sensibility and the linearity of the CCD sensor. These images<br />
were taken at the same time thus in the same lighting and<br />
atmospheric conditions.
210 Nicolas Paparoditis and Olivier Dissard<br />
Figure 3.3.33 Index map of a 60 per cent along track and across track overlap<br />
digital survey on the French town of Amiens.<br />
Multiple views<br />
The acquisition of surveys with multiple views (a particular spot of the<br />
landscape is seen in more than two images) solves many problems. Actually,<br />
provided a sufficient overlapping between images exists, there will always<br />
be one or more couples among the whole set of couples in which a given<br />
point of the landscape is visible and where some of the problems described<br />
previously, e.g. mobile vehicles and non-Lambertian surfaces, will not<br />
appear. In the case of Figure 3.3.33, one landscape point can be seen in<br />
up to 9 images, thus in up to 36 along or across track stereopairs made<br />
out of these 9 images.<br />
Multiple views: merging elementary DSMs processed on all<br />
stereopairs<br />
Means of exploiting this data have been largely treated in recent literature<br />
(Canu et al., 1995; Roux et al., 1998). In most of the techniques, the<br />
general idea is to reconstruct the DSM by an a posteriori merging of all<br />
the elementary DSM calculated on all the possible stereopairs by a voting
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 211<br />
Figure 3.3.34 (a) (b) (c) Three elementary stereopair DSM on the same scene.<br />
One can notice that the hidden areas are not located at the same place; (d) DSM<br />
obtained by merging the three elementary DSMs.<br />
Figure 3.3.35 Zoomed extracts of the DSMs appearing in Figure 3.3.32.<br />
technique (e.g. based on the median value). This merging allows a<br />
densification and reliabilization of the results to a large extent when the<br />
elementary DSM are complementary and when the results are valid in<br />
most of all the DSM. (See Figures 3.3.34 and 3.3.35.)<br />
The stereo matching process requires the use of non-negligible window<br />
sizes if one is looking for sufficiently reliable measures. One knows well<br />
that the morphological quality of the DSM, i.e. the ability to render discontinuities,<br />
slopes, slope breaks, surface microstructures, all of utmost<br />
importance for numerous applications, is directly dependent on the window<br />
size. The larger the windows and the more the ‘high frequencies’ of the<br />
surface are smoothed out, the more the depth discontinuities are delocalized,<br />
and the more the matching of steep slopes is difficult (due to image<br />
deformations). (See Figure 3.3.36.)<br />
One way to reduce the window size is the signal to noise ratio of the<br />
images so as to reduce matching ambiguities. This can be obtained naturally<br />
by having an imaging system of higher quality, or artificially by increasing<br />
the number of observations.
212 Nicolas Paparoditis and Olivier Dissard<br />
Figure 3.3.36 DSM with superimposition of the real roof building edges. One can<br />
see the delocalization of the building depth discontinuities.<br />
Generating the DSM with all the images at the same time: a<br />
multi-image matching process guided from object space<br />
A process based on the concept of direct multi-image matching guided<br />
from object space can be used to process all the images at the same time<br />
to generate the DSM (Figure 3.3.37) (Paparoditis et al., 2000). This process<br />
allows the matching of the images and the construction of the raster DSM,<br />
and the orthoimage all at the same time without having to go through the<br />
processing of all stereopairs and the merging of all elementary DSMs.<br />
The process is based on the following algorithm. For every single node<br />
(x, y) of the DSM raster grid, a correlation profile is constructed gathering<br />
all the correlation scores calculated for each of the possible z through a<br />
plausible interval [zmin, zmax] depending on an a priori knowledge of the<br />
scene (given by a map, a gross DTM, etc.). For each given z, one calculates<br />
in the whole set of images, the hypothetical corresponding (ik , jk )<br />
image coordinates in each image space k (where 0 < k < n and n is the<br />
number of images where this (x, y, z) point is seen). The likelihood of these<br />
hypotheses is given by a direct measurement of the similarity of the set of<br />
windows centred on the (ik , jk ). The ‘estimated’ z value retained for a given<br />
(x, y) is the one for which the optimal value of the correlation profile is<br />
reached. Besides, for this (x, y) and for this estimated z we can calculate<br />
directly the corresponding grey level in the orthoimage from the set of<br />
associated (ik, jk) grey levels by taking for instance the average or the<br />
median of the values.<br />
A multi-image similarity function<br />
How can one generalize the similarity function to a set of image windows?<br />
By defining for example a criterion describing the collinearity dispersion
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
C<br />
C<br />
(x,y)<br />
3D data acquisition from visible images 213<br />
Correlation profile<br />
Figure 3.3.37 Multi-image matching guided from object space.<br />
C<br />
C<br />
of all texture vectors corresponding to all template windows. One can do<br />
this by using this new multi-image texture similarity (MITS) function:<br />
Var <br />
0 MITS (i1, j1, i2, j2, … , in, jn) <br />
n<br />
k1<br />
n<br />
n ,<br />
Var (Vk(i k, jk)) k1<br />
V k (i k ,j k )<br />
(3.20)<br />
where V k (i k , j k ) is the texture vector associated to the window centred on<br />
the pixel (i k , j k ) in image k and Var(V k (i k , j k )) the variance of the texture<br />
vector V k (i k , j k ), i.e. the grey levels inside the window centred on the pixel<br />
(i k , j k ) in image k. Why this correlation function? If the image texture<br />
windows are alike, i.e. the texture vectors are collinear, the similarity score<br />
is maximal.<br />
A radiometric similarity weighting<br />
In view of the great radiometric stability of digital images (which is not<br />
the case of scanned images) and under the assumption that most of the<br />
objects that describe the landscape have Lambertian characteristics, we<br />
z<br />
z max<br />
z<br />
z min
214 Nicolas Paparoditis and Olivier Dissard<br />
Figure 3.3.38 Comparison of homologous windows (3 × 3) found for a given<br />
(x, y) point without and with the radiometric weighting function.<br />
impose an additional constraint on the absolute radiometry of homologous<br />
neighbourhoods, in the shape of a weighting function. This weighting<br />
is necessary. Indeed, the MITS similarity function is ‘centred average’. This<br />
function measures the similarity of the textures of all neighbourhoods but<br />
not their radiometric similarity; this can give rise to mismatches. Our new<br />
similarity function called MITRAS (multi-image texture and radiometric<br />
similarity) can be expressed in the following way:<br />
MITRAS (i 1, j 1, … , i n, j n) <br />
, (3.21)<br />
k<br />
where: Var is the variance, Im (i, j) the grey level of pixel (i, j) in image Im ,<br />
and k a normalizing coefficient. The application of this weighting function<br />
permits a similarity criterion to be obtained which characterizes at<br />
the same time the texture and the radiometric resemblance of the neighbourhoods.<br />
Therefore the MITRAS similarity function is more robust and<br />
discriminating (Figure 3.3.38).<br />
MITS (i 1 , j 1 , … , i n , j n ) exp Var k (I k (i k , j k ))<br />
Multiple view DSM results and window size impact<br />
The increase in the number of observations allows on the one hand a<br />
sizable reduction of matching ambiguities met with in the classical stereoprocessing<br />
and thus to increase the reliability of the process. On the other<br />
hand, it allows, also thanks to the high quality of digital images, sizes for<br />
the correlation windows of 3 × 3 to be used (Figure 3.3.39). The results<br />
show clearly a dense DSM, an extremely good morphological rendering:<br />
all relief slopes, slope breaks, discontinuities and microstructures, and a
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 3.3.39 DSM generated with 4 images and 3 × 3 windows.<br />
Figure 3.3.40 DSM generated with 4 images and 5 × 5 windows.
216 Nicolas Paparoditis and Olivier Dissard<br />
good localization of the discontinuities (especially the building edges).<br />
However, the very small size of the windows is at the origin of false<br />
matches in the very homogeneous areas that do not appear with a 5 × 5<br />
window (Figure 3.3.40). To the contrary, a 5 × 5 window has a less accurate<br />
rendering of microstructures and discontinuities.<br />
Working from object space offers a good number of advantages. One<br />
can work in a transparent way on images of different resolutions, and the<br />
parameters are expressed in a metric form. Also we avoid the blind resampling<br />
of all the 3D samples corresponding to all image matches in order<br />
to generate a regular grid DSM in object space. Indeed, these samples<br />
when the matching process is guided from image space follow, although<br />
their distribution is regular in image space, an irregular spatial distribution<br />
in object space. Finally, its major advantage is that it allows, on the<br />
Figure 3.3.41 Orthoimage corresponding to the DSM of Figure 3.3.40. Images<br />
with 40 cm ground pixel size from the digital camera of IGN-F on<br />
the French town of Le Mans with a 50 per cent across and along<br />
track overlap.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 217<br />
Figure 3.3.42 2,000 × 2,000 Raw DSM obtained with our process on 4 images<br />
and 3 × 3 windows.<br />
one hand N images to be treated in a natural and transparent way, and<br />
on the other hand, the DSM and the corresponding orthoimage to be<br />
constructed at the same time.<br />
Matching self-evaluation and automatic filtering of false<br />
matches<br />
As expected, the multi-image correlation score does not supply a probability,<br />
but a good self-indicator of the measurement reliability. The results<br />
(in Figure 3.3.42 and Figure 3.3.43) show that DSM aberrant points have<br />
much weaker correlation scores.<br />
A simple experiment of analysis of the correlation score distributions<br />
shows this very clearly. Let us consider real and aberrant score distributions.<br />
We call real scores those found with our process, and aberrant scores<br />
those obtained by the same process with the proviso that one deliberately<br />
stands within an altimetric search interval of the same amplitude but<br />
containing no relief. One can observe that the two distributions are
218 Nicolas Paparoditis and Olivier Dissard<br />
Figure 3.3.43 Correlation scores corresponding to Figure 3.3.42.<br />
0.000<br />
0.230<br />
0.460<br />
0.690<br />
0.920<br />
1.150<br />
1.380<br />
1.610<br />
1.840<br />
2.070<br />
2.300<br />
MITRAS on 4 images<br />
2.530<br />
2.760<br />
2.990<br />
3.220<br />
3.450<br />
3.680<br />
3.910<br />
Figure 3.3.44 Correlation score histograms.<br />
real matches with 5×5 window<br />
real matches with 3×3 window<br />
false matches<br />
but slightly mixed (Figure 3.3.42). From an inspection of the curves, it<br />
can be said that on this scene the points having correlation scores above<br />
2.9 are almost 100 per cent sure. (See Figure 3.3.44.)<br />
The relative distribution separation in the case of multi-image matching<br />
allows a valid criterion for false points filtering to be defined. It can be<br />
noticed that the greater the number of images, the greater is the separation<br />
of the two distributions. In the case of stereo-processing, the mixing
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3D data acquisition from visible images 219<br />
Figure 3.3.45 DSM of Figure 3.3.42 filtered with a 2.9 threshold on the<br />
correlation scores of Figure 3.3.42.<br />
of the distributions is considerable. This explains the difficulty and the<br />
impossibility of defining a reliability criterion and a satisfying rejection<br />
threshold on the correlation score values. A low threshold retains a great<br />
number of aberrant points, while a high one rejects a very great number<br />
of valid points. (See Figure 3.3.45.)<br />
This matching process tends voluntarily to determine solely the landscape<br />
points seen in a rather similar way on all the images. As a result,<br />
the filtering process will remove all other points and leave missing areas<br />
inside the DSM. For example surface points that are only seen in a subset<br />
of images will be removed. Nevertheless, we have here the first step of a<br />
general matching strategy that can be to complete in a progressive and<br />
hierarchical way the DSM starting from the most reliable measures.
220 Nicolas Paparoditis and Olivier Dissard<br />
3.3.10 Generating a DTM from a DEM: removing buildings,<br />
bridges, vegetation, etc.<br />
Nicolas Paparoditis<br />
In the case of dense urban areas and extended areas of vegetation, even if the<br />
stereo-process is carried out on SPOT like low-resolution images, the<br />
measured elevation samples will describe the rooftops and the canopy<br />
surface. The only proper way of generating a DTM from a DEM is by removing<br />
from the DEM all the samples corresponding to 3D raised structures, i.e.<br />
vegetation, buildings, etc. and fitting surfaces through the holes. This process<br />
requires a 3D analysis of the DEM, which can be combined with an image<br />
analysis to detect and recognize all these cartographic features.<br />
References<br />
Baillard C., Dissard O. (2000) A stereo matching algorithm for urban digital<br />
models, Photogrammetric Engineering and Remote Sensing, vol. 66, no. 9,<br />
September, pp. 1119–1128.<br />
Baillard C., Maître H. (1999) 3D reconstruction of urban scenes from aerial stereo<br />
imagery: a focusing strategy, Computer Vision and Image Understanding, vol. 76,<br />
no. 3, December, pp. 244–258.<br />
Bertero M., Poggio T., Torre V. (1988) Ill-posed problems in early vision, IEEE<br />
Proceedings, vol. 76, no. 8, pp. 869–889.<br />
Canu D., Ayache N., Sirat J.A. (1995) Accurate and robust stereovision with a<br />
large number of aerial images, SPIE (Paris), vol. 2579, pp. 152–160.<br />
Cord M., Paparoditis N., Jordan M. (1998) Dense, reliable and depth discontinuity<br />
preserving DEM computation from H.R.V. images, Proceedings of ISPRS Commission<br />
II Symposium, Data Integration: Systems and Techniques, Cambridge,<br />
England, July, vol. 32, no. 2, pp. 49–56.<br />
Gabet L., Giraudon G., Renouard L. (1994) Construction automatique de modèles<br />
numériques de terrain à haute résolution, Bulletin SFPT, no. 135, pp. 9–25.<br />
Grimson W. (1983) An implementation of a computational theory of visual surface<br />
interpretation, Computer Vision Graphics Image Processing, no. 22, pp. 39–69.<br />
March R. (1989) A regularization model for stereo vision with controlled continuity.<br />
Pattern Recognition Letters, no. 10, pp. 259–263.<br />
Okotumi M., Kanade T. (1993) Multiple Baseline Stereo, IEEE Transactions on<br />
Pattern Analysis and Machine Intelligence, vol. 15, no. 4, April, pp. 353–363.<br />
Paparoditis N., Cord M., Jordan M., Cocquerez J.P. (1998) Building detection and<br />
reconstruction from mid to high resolution images, Computer Vision and Image<br />
Understanding, 72(2), November, pp. 122–142.<br />
Roux M., Leloglu U.M., Maître H. (1998) Dense urban DEM with three or more<br />
high resolution aerial images, ISPRS Symposium on GIS, Between Vision and<br />
Applications, Stuttgart, vol. 32, no. 4, September, pp. 347–352.<br />
Tikhonov A.N. (1963) Solution of incorrectly formulated problems and the regularization<br />
method, Soviet Mathematical Document, no. 4, pp. 1035–1038.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3.4 FROM THE DIGITAL SURFACE MODEL (DSM) TO<br />
THE DIGITAL TERRAIN MODEL (DTM)<br />
Olivier Jamet<br />
As soon as one is interested in the topographic data production with a<br />
rms accuracy in altitude of the order of a metre or less, the derivation of<br />
a DTM from the calculated DSM, by manual edition or automatic levelling<br />
of the raised structures, is an indispensable step. The semi-automatic<br />
processes permitting one to alleviate the edition of the DSM can be classified<br />
in three categories.<br />
3.4.1 Filtering of the DSM<br />
From the DSM to the DTM 221<br />
Techniques of filtering are based on the hypothesis that elements of the<br />
raised structures constitute connex zones of limited area and presenting<br />
strong contrasts of elevation with their environment. The method most<br />
often used is issued from mathematics morphology: the DTM is obtained<br />
by morphological opening of the DSM by a structuring element defined<br />
by its size and its geometry (parameters of the filter). This method is<br />
perfectly rigorous only in a plane land. Indeed it doesn’t preserve the curvature<br />
of the land. The quality of the DTM produced will be therefore worse<br />
if the land is hilly and if the structuring element is of large size. Besides,<br />
the abrupt irregularities of the land leads to artefacts that can become very<br />
visible for openings of large size, as in the case of the elimination of<br />
building envelopes in a metric resolution DSM, for example. These artefacts<br />
can be accentuated by the noise of the DSM, and more precisely by<br />
the presence of local parasitic minima. A filtering of these local minima,<br />
for example by an applied morphological closing of the DSM, will then<br />
improve the results.<br />
A variant of this method of levelling of the raised structures consists in<br />
applying openings of increasing successive sizes with structuring elements<br />
of variable shape up to the wanted size. If this method doesn’t preserve<br />
the accuracy of the DTM better in the presence of strong curvatures, it<br />
has the advantage of producing a smoother result.<br />
3.4.2 Segmentation and elimination of the raised structures<br />
An alternative to the simple filtering consists in using techniques of shape<br />
recognition to circumscribe the extension of the raised structures. The<br />
DTM is then calculated by replacing, on the extension of the detected<br />
raised structures, the initial elevation values by interpolated values. This<br />
segmentation of the raised structures can be done by a process of the DSM<br />
only, or by analysing jointly an orthophotography of the same zone.
222 Olivier Jamet<br />
3.4.2.1 Segmentation of the DSM<br />
All tools of segmentation used in picture analysis can obviously be used<br />
for the analysis of the DSM, requiring a definition of the criteria characterizing<br />
the raised structures in relation to the ground.<br />
The simplest methods are based on the examination of slopes or curvatures<br />
of the DSM for a detection analogous to a detection of contour in<br />
an image. The selection of the raised structures in the extracted contours<br />
is made by imposing that contours are closed and that the value of the<br />
slope on contours is greater than a minimal value. A supplementary<br />
constraint on the size of the detected zones can also be applied. These<br />
techniques are not recommended. They will always be extremely sensitive<br />
to the accuracy of the DSM at the edge of the raised structures, though<br />
this accuracy is impossible to guarantee (either because of qualities of the<br />
picture having served to the extraction, or because of the hidden parts in<br />
the case of a matching on a couple of pictures).<br />
It appears therefore preferable, if one uses only elementary operations<br />
of image process, to use the method described in §3.4.1. A simple threshold<br />
on the difference between the DSM and the DTM produced by morphological<br />
opening will indeed give a segmentation of the raised structures<br />
a priori more reliable than techniques of contour extraction – this<br />
segmentation permitting the calculation of a new DTM by initial data<br />
interpolation.<br />
To palliate the fragility of methods based on contour extraction on the<br />
DSM and the shortcomings of methods based on mathematical morphology,<br />
which will fail systematically in cases of extended raised structures<br />
(islets of buildings in urban zone, forests, etc.) and hilly landscapes, some<br />
more elaborate methods are made the object of research. In Baillard (1997),<br />
the raised structures is defined by a Markovian model in which the labelling<br />
of a region as raised structures is a function of relations of dependence<br />
with its neighbouring regions (e.g. ‘a zone higher than a neighbouring zone<br />
recognized as belonging to the raised structures necessarily belongs to<br />
the raised structures’). Methods of segmentation by area growth, with a<br />
criterion of homogeneity based on differences of altitude between neighbouring<br />
points are proposed also in Masaharu and Hasegawa (2000), and<br />
Gamba and Casella (2000) for example. Other authors use some cartographic<br />
data to force the segmentation (e.g. Horigushi et al. (2000)). (See<br />
Figure 3.4.1, colour section.)<br />
3.4.2.2 Segmentation of an orthophotography superimposable<br />
on the DSM<br />
The interpretation of photographic pictures of the studied zone, if not alone<br />
able to distinguish the raised structures without ambiguity, can improve the<br />
reliability of processes of segmentation of the DSM. In particular, specific
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
detectors (recognition of the structure, vegetation) applied on an orthophotography<br />
superimposable on the DSM will provide complementary masks<br />
combinable with the segmentation of the DSM. Dissard applies this technique,<br />
for example, for the detection of forest zones.<br />
3.4.3 Selection of ground points in the DSM<br />
From the DSM to the DTM 223<br />
Figure 3.4.1 Images from the digital camera of IGN-F (1997) on Le Mans<br />
(France). The outstanding signal/noise ratio allows one to get<br />
excellent matching results on the automatic correlation process,<br />
particularly on homogeneous or shadowy zones where normal<br />
digitized images always produce many difficulties. Benefit is taken<br />
from this situation, by using small correlation windows (5 × 5 and<br />
even 3 × 3), which improves the planimetric accuracy.<br />
A third approach consists in selecting, by an ad hoc procedure, a set of<br />
points situated on the ground in the DSM, and rebuilding the DTM by<br />
interpolation of this set of points. This technique has the advantage of<br />
avoiding the problem of an exact segmentation of the extensions of the
224 Olivier Jamet<br />
raised structures. It can, however, be used only in the individual cases<br />
where the strategy of selection of ground points is very reliable (the sensitivity<br />
to errors being as large as the density of the selected seeding is low).<br />
It is, for example, the case in urban zones, when the land varies little.<br />
One can be content with a scattered set of points, which one can choose,<br />
for example, on the road network (if a cartography is available, or while<br />
calculating an automatic extraction).<br />
In the absence of other information, it is also possible to use procedures<br />
of blind selection of local minima in an analysis procedure of the DSM<br />
by mobile window, the window of analysis being chosen of sufficient size<br />
not to be ever entirely covered by the raised structures. This method,<br />
however, is rigorous only in flat landscapes and is extremely sensitive to<br />
the noises of the DSM.<br />
3.4.4 Conclusion<br />
These methods cannot be considered as perfectly operative, however.<br />
Only the simplest algorithms (as algorithms of filtering that are most<br />
current) are generally present on digital photogrammetric workstations.<br />
They will give most of the time good results on landscapes of moderate<br />
slope and when the raised structures is constituted of connex parts of weak<br />
extension. They remain therefore precious from the moment the number<br />
of these connex parts is high, because they will allow one to save a lot of<br />
editing time.<br />
They are, however, very sensitive to the quality of the calculated DSM,<br />
and in particular to the slope of the DSM at the borders of the raised<br />
structures (that must be the strongest possible). They can also produce<br />
some unexpected results on ambiguous superstructures such as bridges,<br />
which can either be considered as belonging to the raised structures or<br />
not. They must not therefore be used without supervision.<br />
References<br />
Baillard C. (1997) Analyse d’images aériennes stéréoscopiques pour la restitution<br />
3D des milieux urbains: Détection et caractérisation du sursol. Ph.D. Thesis,<br />
École Nationale Supérieure des Télécommunications, Paris.<br />
Horigushi S., Ozawa S., Nagai S., Sugiyama K. (2000) Reconstructing road and<br />
block from DEM in urban area, Proc. ISPRS Congress 2000, International Archives<br />
of <strong>Photogrammetry</strong> and Remote Sensing, vol. XXXIII, Part B3, pp. 413–420.<br />
Gamba P., Casella V. (2000) Model independent object extraction from digital<br />
surface models, Proc. ISPRS Congress 2000, International Archives of <strong>Photogrammetry</strong><br />
and Remote Sensing, vol. XXXIII, Part B3, pp. 312–319.<br />
Masaharu H., Hasegawa H. (2000) Three-dimensional city modelling from laser<br />
scanner data by extracting building polygons using region segmentation method,<br />
Proc. ISPRS Congress 2000, International Archives of <strong>Photogrammetry</strong> and<br />
Remote Sensing, vol. XXXIII, Part B3, pp. 556–562.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3.5 DSM RECONSTRUCTION<br />
Grégoire Maillet, Patrick Julien, Nicolas Paparoditis<br />
3.5.1 <strong>Digital</strong> elevation models under TIN shape<br />
Grégoire Maillet<br />
DSM reconstruction 225<br />
The natural form of the DEM is raster (a regular grid of altitudes), either<br />
acquired by correlation or by laser scanning. But, given their extremely<br />
heterogeneous nature and the presence of discontinuities, this format is<br />
neither adapted to their storage nor to their manipulation. The construction<br />
of a TIN (for triangulated irregular network) presents the two favourable<br />
features to reduce greatly the volume of data (a reduction of around 90 per<br />
cent for an urban DEM) and to be more effective for all applications of<br />
visualizations [1] (Figure 3.5.1.5). Besides, this construction is a first raw<br />
data interpretation that may facilitate the analysis of the scene [2].<br />
Contrary to the general problems of surface triangulation, the raster origin<br />
of data makes it possible here to work in 2.5D. Most of the existing<br />
methods are based on the recursive insertion of points in a planimetric triangulation<br />
up to the satisfaction of a criterion on the number of points or<br />
on the quality of the approximation [3]. The problem can be decomposed<br />
therefore into two independent sub-problems: the way to triangulate points<br />
and the criterion of the choice of the points to insert.<br />
A triangulation by a simple splitting into three of the triangle to the<br />
vertical of the point to insert leads inevitably to triangles of very acute shape<br />
that are very difficult to manipulate. The criterion of Delaunay allows the<br />
appearance of these problems to be minimized, and there are very efficient<br />
implementations of insertion of points in a Delaunay triangulation [4][3].<br />
Nevertheless, the lack of taking into account the altitude can lead to optimal<br />
triangulations in planimetry, but disjointed in 3D (Figure 3.5.1.1).<br />
Different methods exist to take into account the altimetric consistency<br />
of the triangulation [3] (Figures 3.5.1.2 and 3.5.1.3), but the altimetric<br />
improvement is often performed to the detriment of the planimetric quality<br />
of the triangulation. The choice of the points to insert is based in general<br />
on a measure of importance, the simplest of these measures being the local<br />
error on the considered point, that is to say the gap in Z between the TIN<br />
and the raster model. It is a simple and fast measure [3] but one that<br />
remains quite sensitive to the noise of the DEM and provides not very<br />
meaningful results to the neighbourhood of altimetric discontinuities.<br />
A more robust measure consists in evaluating the volume added by the<br />
insertion of a point while taking the volume between the point to insert<br />
and the triangle that is on its vertical. This measure will give some more<br />
meaningful results in the case of light delocalizations on the sides of the<br />
buildings (Figure 3.5.1.4).<br />
At the time of the construction of a TIN by iterative insertion many<br />
superfluous points are inserted. They are judged optimal to one given
226 Grégoire Maillet et al.<br />
Figure 3.5.1.1 On the left the triangulation of a building and its perspective<br />
representation with the criterion of Delaunay, on the right a<br />
triangulation taking into account the altitudes.<br />
Figure 3.5.1.2 At the time of the insertion of a point there is division of the<br />
triangle into three. For each neighbouring quadrilateral the two<br />
diagonals are tested. And the solution creating fewer errors in<br />
Z is preserved.<br />
instant and permit a convergence toward a good solution but become<br />
redundant in the final result. The simplification of surface is therefore a<br />
complementary and necessary approach for the construction of an optimal<br />
TIN, that is to say describing in the best way the DEM with a minimum<br />
of points. This simplification can be performed in a way similar to the<br />
insertion: with a method of suppression of points in the triangulation and<br />
a measure of the importance of points. But the simplification of TIN is a<br />
problem very often met in image synthesis and thus specific methods have<br />
been invented. The most frequent approach is not based on the suppression<br />
of points, but on the contraction of a pair of points. That is to say
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 3.5.1.3 For every triangle one looks for four characteristic points (one in<br />
the triangle and one on each side). According to the result of this<br />
research one applies one of the five possible divisions.<br />
C<br />
A<br />
P<br />
H<br />
B<br />
Figure 3.5.1.4 For a point P on a side of roof slightly delocalized the local error<br />
PH is large whereas the volume of the ABCP tetrahedron is small.
228 Grégoire Maillet et al.<br />
Figure 3.5.1.5 (a) urban DEM obtained by correlation; (b) transformation into<br />
a TIN containing 5 per cent from initial points; (c) perspective<br />
visualization of the TIN.<br />
the replacement of a couple of points by a new point minimizing the<br />
committed error [5].<br />
References<br />
[1] Haala N., Brenner C. Virtual city models from laser altimeter and 2D map<br />
data. Photogrammetric Engineering and Remote Sensing, vol. 65, no. 7, July<br />
1999, pp. 787–795.<br />
[2] Maas H.-G., Vosselman G. Two algorithms for extraction building models from<br />
raw laser altimetry data. ISPRS Journal of <strong>Photogrammetry</strong> and Remote<br />
Sensing, vol. 54, nos. 2–3, July 1999, pp. 153–163.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
[3] Garland M., Heckbert P.S. Fast polygonal approximation of terrains and height<br />
fields. Technical Report CMU-CS-95–181, Comp. Sci. Dept, Carnegie Mellon<br />
University, September 1995, pp. 1–37.<br />
http://www.cs.cmu.edu/~garland/scape<br />
[4] Devillers O. Improved incremental randomized Delaunay triangulation. INRIA<br />
Research Report no. 3298, November 1997, pp. 1–21.<br />
[5] Garland M., Heckbert P.S. Surfig. Simplification using quadric error metrics.<br />
SIGGRAPH 97 Proc., August 1997, pp. 209–216.<br />
http://www.cs.cmu.edu/~garland/quadrics<br />
3.5.2 Constitution of a DTM in the shape of a regular grid<br />
of altitudes, from contour lines in vector mode<br />
Patrick Julien<br />
DSM reconstruction 229<br />
3.5.2 Position of the problem. Notion of topographic surface<br />
Suppose one is given a file containing a system of contour lines describing<br />
the relief of a fixed geographical zone. To simplify, we call a contour line<br />
what is in fact a section of a contour line, either closed, or stopped on<br />
the edge of the zone. Every contour is assumed to be represented in vector<br />
mode, i.e. described by vertices of a polygonal line, or by an ordered<br />
sequence of points, say {pij (xij, yij, zj); i 1, . . ., n(j)} for the contour<br />
1<br />
M<br />
l<br />
2<br />
1<br />
p ij<br />
2 c<br />
N<br />
1<br />
p i+l, j<br />
m cl<br />
Figure 3.5.2.1 Contour lines in vector mode.<br />
h<br />
p n(j), j<br />
h
230 Grégoire Maillet et al.<br />
line placed to the rank j of the file; the coordinate z is the altitude, and<br />
the x, y coordinates are relative to a cartographic representation. It is<br />
important to point out that the contour line is not only the finite family<br />
of points p ij, but the union of segments [p ij, p i1,j]; it is therefore a continuous<br />
series, whose family p ij is one of the possible representations; in<br />
particular, one is allowed to subdivide any segment [p ij , p i1,j ] into smaller<br />
segments, which doesn’t change the shape of the curve. (See Figure 3.5.2.1.)<br />
Now, one wishes to ‘interpolate’ in these contour lines a grid of points<br />
with a grid interval h:<br />
{m cl (x c, y l, z cl); c 1, . . ., N, l 1, . . ., M}, (3.22)<br />
where x c ch, y l lh, regular in the x, y coordinates system. By ‘interpolating’,<br />
we mean that every point must be situated on ‘the’ surface<br />
represented by the contour lines, which surface is in principle well defined<br />
by the rule of surveying these contours, according to which rule their plot<br />
should be such that the user can always consider the slope between two<br />
regular or intermediate curves as regular.<br />
For example, the rule above means that if the user wants to draw a new<br />
contour line situated at half height of two successive ones, he must place<br />
it at an equal horizontal distance from the first ones, for if it were not the<br />
case the slope would be steepest on one side of the new contour than on<br />
the other, and therefore non-regular; more generally any intermediate<br />
contour must be plotted between the two neighbouring contours, with<br />
horizontal distances in ratio to vertical distances. In a more precise way,<br />
we will interpret the rule while saying that, between two successive contours<br />
and along a steepest slope line, the scalar value of the slope can be considered<br />
constant. One so rejoins the rule according to which contour lines<br />
must permit us to determine, with a sufficient precision, the elevations of<br />
any ground point by simple linear interpolation.<br />
We insist on the fact that it is necessary to admit that these rules give<br />
place, for several users, to negligible interpretation differences, in other<br />
words that contour lines suggest to all users the same ‘topographic’ surface<br />
(simplified model of the real ground surface), obtained in practice by the<br />
rule of the linear interpolation next along the line of steepest slope; it is<br />
on this surface that the grid points m cl must lie.<br />
So the interpolation of points m cl is transformed into the plotting of the<br />
steepest slope lines, which are the orthogonal lines to contours (Figure<br />
3.5.2.2). However, the tracing of these orthogonal lines, that essentially<br />
consists in joining between each other bootjacks placed on curves, require,<br />
as soon as orthogonal lines are a little wavy, a certain manual ability. One<br />
guesses that the algorithmic transposition of this ability is delicate, and<br />
the methods presented below only achieve it in an approximate way.<br />
We are going to describe here methods of approximation by triangular
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
irregular network surface, by thin plate spline surface, and by elastic grid<br />
surface.<br />
3.5.2.2 Approximation of the topographic surface by a<br />
triangular irregular network surface (TIN surface)<br />
DSM reconstruction 231<br />
Figure 3.5.2.2 Handmade drawing of the steepest slope lines, orthogonal to<br />
contours.<br />
Definitions relative to triangulations<br />
Let’s first specify the vocabulary used. A triangulation of a domain D of<br />
the plane is a finite family of triangles of which union is D, of non-null<br />
areas, of which any two have in common either nothing, or a vertex, or<br />
a side. A triangulation is called Delaunay’s if the open circle (i.e. a circle<br />
without its edge) circumscribed to any triangle contains no other triangle<br />
vertex. One shows that, given a finite family {mi} of points of the plane,<br />
there exists a triangulation of Delaunay, and algorithms to construct it,<br />
whose vertices of triangles are exactly the points mi; the triangulation is<br />
unique if there are never four cocyclic points mi, mj, mh, mk. A triangulated<br />
surface is a connex surface, union of a family of (plane) triangles of<br />
the space, of non-null areas, of which any two have in common either<br />
nothing, or a vertex, or a side.
232 Grégoire Maillet et al.<br />
Conditions to be observed by a TIN surface approximating the<br />
topographic surface<br />
We have now to approximate the topographic surface with a triangulated<br />
surface fitted on contour lines. The searched-for surface here is with<br />
irregular faces, and the horizontal positions of triangle vertices are not a<br />
priori known; this problem is therefore distinct from fitting a surface whose<br />
horizontal positions of vertices would form a fixed in advance network,<br />
a problem in which only the altitudes of triangle vertices are unknown.<br />
It seems natural that triangles’ vertices are points taken on contour lines,<br />
because there doesn’t appear a simple way to define other points for this<br />
use. Besides, every triangle must be an acceptable approximation of the<br />
topographic surface, which requires that its three sides are close to this<br />
surface. That in particular notably forbids that a triangle edge ‘crosses’ a<br />
contour (in the sense that their horizontal projections intersect each other),<br />
because, in general, such an edge no longer lies on the topographic surface;<br />
for example, on Figure 3.5.2.3, the AB side crossing the contour (z + e)<br />
does not lie on the topographic surface whose profile is AIB. In particular,<br />
a triangle cannot have its vertices on three contours of different<br />
altitudes, otherwise one of its sides would ‘cross’ a contour (ABC triangle).<br />
A triangle must not have its three vertices on the same curve, otherwise<br />
it would be horizontal and would not lie on the topographic surface (DEF<br />
triangle). It is therefore necessary that every triangle has its vertices on<br />
two consecutive contours; in addition, the side defined by the two vertices<br />
placed on the same contour must be confounded with the contour (GHJ<br />
triangle) for, if it were not the case, this side could not lie on the surface.<br />
z<br />
E<br />
z+e<br />
D<br />
H<br />
F<br />
z+2e<br />
J<br />
C<br />
G<br />
B<br />
I<br />
Figure 3.5.2.3 ABC is an unauthorized triangle because it touches three contours;<br />
DEF is unauthorized because it touches only one contour; GHJ is<br />
an authorized triangle, touching two consecutive contours, with<br />
one side along a contour.<br />
A
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
It is necessary to also observe, for example on the GHJ triangle, that<br />
the smaller the HJ side in relation to the sides GH and GJ, the better is<br />
the approximation of the topographic surface; it is therefore a priori preferable<br />
that sides bordering the curves be short.<br />
It is finally necessary that triangles overlap the whole zone described by<br />
contour lines, so that every point of any contour line is on a side of a<br />
triangle; and that above every horizontal position (x, y) there is only one<br />
triangle. In these conditions, the horizontal projections of triangles constitute<br />
a plane triangulation. Once the future triangle vertices (X i, Y i, Z i) are<br />
chosen, one will therefore get a triangulated surface by constructing a<br />
plane triangulation whose vertices of triangles are points (X i, Y i).<br />
Construction of a triangulated surface<br />
Among all possible plane triangulations, that of Delaunay seems a reasonable<br />
choice, because its geometric properties are known, unlike a<br />
triangulation that would be constructed with the help of heuristics.<br />
It remains to choose the points devoted to be the triangles’ vertices. It<br />
is necessary to take at least all points pi,j initially defining the contour<br />
lines (except possibly those aligned with their predecessor pi1,j and their<br />
successor pi1,j , which doesn’t modify the tracing of the contour). However,<br />
if the distance between two consecutive points pi,j, pi1,j is large in relation<br />
to their horizontal distances d1, d2 to neighbouring curves, the triangulation<br />
of Delaunay risks the construction of triangles crossing a curve,<br />
therefore forbidden.<br />
For example from the four points p, p′, q, r of Figures 3.5.2.4(a) and<br />
(b), one can construct either the authorized pair of triangles {pp′q, pp′r},<br />
or the pair {pqr, p′qr}, forbidden because the side pq ‘crosses’ a contour.<br />
In the case where angle (q) angle (r) < < angle (p) angle (p′) (Figure<br />
3.5.2.4(a)), the pair of triangles is compatible with the condition of<br />
Delaunay because neither r is in the pp′q circle, nor q in the pp′r circle,<br />
whereas the {pqr, p′qr} pair is incompatible since p is in the p′qr circle<br />
and p′ in the pqr circle. In the opposite case where angle (p) angle (p′)<br />
< < angle (q) angle (r) (Figure 3.5.2.4(b)), the pair {pqr, p′qr} is Delaunay<br />
compatible, but not the pair {pp′q, pp′r}.<br />
So in order that the Delaunay triangulation constructs the acceptable pair,<br />
it is necessary to be in the case: angle (q) angle (r) < ; as angle (q) < 2<br />
Arctan (d/2d1 ) and angle (r) < 2 Arctan (d/2d2 ), it is sufficient that:<br />
Arctan d<br />
2d 1 Arctan d <br />
2d <<br />
2 2 Arctan d<br />
2d 1 Arctan 2d1 d ,<br />
i.e. d < 2 √d 1 d 2 .<br />
DSM reconstruction 233<br />
(3.23)
234 Grégoire Maillet et al.<br />
d 1<br />
d 2<br />
p<br />
q q<br />
r<br />
d<br />
p′<br />
Figure 3.5.2.4 (a) (left) Topographically authorized pp′q and pp′r triangles are<br />
selected by Delaunay triangulation; (right) Topographically<br />
unauthorized triangles pqr and p′qr are rejected by Delaunay<br />
triangulation.<br />
d 1<br />
d 2<br />
p<br />
q<br />
r<br />
p′<br />
d d<br />
Figure 3.5.2.4 (b) (left) Triangles pq′q and pp′r are topographically authorized,<br />
but rejected by Delaunay triangulation; (right) Triangles pqr and<br />
p′qr are topographically unauthorized, but selected by Delaunay<br />
triangulation.<br />
If d does not verify this condition, additional points between p and p′ are<br />
necessary, at mutual distances d′
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
p<br />
p′<br />
p′′<br />
q<br />
DSM reconstruction 235<br />
Figure 3.5.2.5 Topographically authorized triangle pp′q is rejected by Delaunay<br />
triangulation; instead the horizontal unauthorized triangle pp′p′′ is<br />
selected.<br />
r<br />
q m<br />
q 3<br />
q2<br />
q 1<br />
q<br />
Figure 3.5.2.6 From the horizontal triangles (in grey) selected by Delaunay<br />
triangulation, a polygonal line qq 1 q 2 ... q m is defined, with each<br />
q i in the middle of an edge, and: Altitude (q) > altitude (q 1 )><br />
altitude (q 2)>...>altitude (q m) > altitude (r).<br />
One constructs then a new Delaunay triangulation (Figure 3.5.2.7), if<br />
necessary by adding some intermediate points on segments [q 1, q 2],<br />
[q 2, q 3], . . ., [q m1, q m], to guarantee that no triangle rides the line qr. Then<br />
the TIN surface thus constructed is a relatively satisfactory approximation<br />
of the topographic surface.<br />
Going from TIN surfaces to the regular grid of altitudes<br />
Once the surface with triangular faces is obtained, it remains to determine<br />
in which triangle of the triangulation each mcl point is projected, to form
236 Grégoire Maillet et al.<br />
Figure 3.5.2.7 A part of the improved Delaunay triangulation (including the<br />
additional vertices q q 1 q 2 ... q m) with topographically authorized<br />
triangles.<br />
the equation z ax by d of the plane of this triangle, and to calculate<br />
the elevation z cl ax c by l d.<br />
3.5.2.3 Approximation of the topographic surface by a thin<br />
plate spline surface<br />
Definition<br />
Thin plate spline surfaces answer the following problem of interpolation:<br />
given a sample of n points of the plane mi (xi, yi) and their elevation<br />
zi, find a function Z(x, y) so that Z(xi, yi) zi for i 1, . . ., n and that<br />
the integral<br />
K(Z) RR<br />
(3.24)<br />
be minimal; Z′′<br />
xx , Z′′<br />
xy , Z′′<br />
yy designate partial second derivatives.<br />
Thin plate splines also answer the adjustment problem: besides, given n<br />
weight values i > 0, find a function Z(x, y) such that the quantity:<br />
E(Z) K(Z) n<br />
Z′′<br />
xx(x, y) 2 2Z′′<br />
xy(x, y) 2 Z′′<br />
yy(x, y) 2 dx dy<br />
i1<br />
i Z(x i, y i) z i 2<br />
(3.25)<br />
be minimal.<br />
J. Duchon showed (1976) that each of these problems admits a solution<br />
and only one (under the condition – always true – that the sample<br />
counts at least three non-aligned points). He also characterized solutions<br />
as functions in the shape:
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
with<br />
where<br />
(3.26)<br />
Note that Z(x, y) is defined by continuity at (x i, y i) points since (r) tends<br />
toward 0 when r tends toward 0.<br />
More precisely, the solution of the interpolation problem is the function<br />
Z(x, y) of the shape Eqn (3.26) such that Z(x j, y j) Z j(j 1 . . . n); i.e.<br />
whose n3 coefficients a 1, . . ., a n,b 0, b 1, b 2 verify the n3 equations:<br />
with<br />
(3.27)<br />
The solution of the adjustment problem weighted by the family { i }<br />
is the function Z (x, y) of shape (3.26) such that (8a j/ j) Z (x j, y j) <br />
z j (j 1 . . . n), i.e. whose n3 coefficients verify the n3 equations:<br />
with<br />
Z(x, y) i<br />
i<br />
i<br />
i<br />
8aj<br />
i<br />
j<br />
a i 0, i<br />
a i 0, i<br />
i<br />
a i 0, i<br />
a ir i (x, y) b 0 b 1 x b 2 y ,<br />
a i x i 0, i<br />
r i(x, y) √(x x i) 2 (y y i) 2 , (r) r 2 ln r.<br />
a i r i (x j , y j ) b 0 b 1 x j b 2 y j z j ( j 1,… , n),<br />
a i x i 0, i<br />
ai ri (xj , yj ) b0 b1xj b2yj zj (j 1, … , n), (3.28)<br />
a i x i 0, i<br />
a i y i 0,<br />
a i y i 0.<br />
a i y i 0.<br />
DSM reconstruction 237<br />
Moreover, these equations clearly show the interpolation problem as the<br />
limit of the adjustment problem when weights j become arbitrarily large.<br />
A surface Z(x, y) of the shape (3.26) has been called a thin plate spline<br />
by reason of the physical interpretation of the integral K(Z), roughly proportional<br />
to the energy of bending of a thin plate of equation z Z(x, y).<br />
Geometric significance of the integral K(Z)<br />
Geometrically, the integral K(Z) represents the ‘global’ curvature of the<br />
surface Z(x, y). Indeed a Z function, supposed twice differentiable at
238 Grégoire Maillet et al.<br />
(x, y), is approximated at the points (xr, ys) near (x, y) by its Taylor<br />
polynomial of degree 2:<br />
Z(xr, ys) Px,y(r, s) (3.29)<br />
where Px,y (r, s) Z(x, y) Z′ xr Z′ ys is the tangent plane in (x, y).<br />
One sees that the surface is all the more distant from its tangent plane<br />
(and so all the more concave), as Z′′ xx, Z′′ xy Z′′ yx, Z′′ yy are more distant<br />
2 2 2 2 2 from 0, or as Z′′ xx , Z′′ xy , Z′′ yy are larger; so the quantity Z′′ xx 2Z′′ xy <br />
2 Z′′ yy measures the curvature of the Z surface in every point.<br />
1<br />
2 Z′′ xxr 2 Z′′ xyrs Z′′ yxrs Z′′ yys2, REMARK<br />
2 2 2 The quantity Z′′ xx 2Z′′ xy Z′′ yy has the advantage over similar measures<br />
such as |Z′′ xx| |Z′′ xy| |Z′′ yy| or sup(|Z′′ xx|, |Z′′ xy|, |Z′′ yy|) of being a characteristic<br />
of the surface, invariant by a rotation of the orthonormal reference<br />
frame, i.e. by change of variable (x, y) (X cos a Y sin a, X sin a Y<br />
cos a). For the U parametrization of the Z surface defined by U(X, Y) <br />
Z(X cos a Y sin a, X sin a Y cos a) Z(x, y), one has:<br />
U′′ XX(X,Y) <br />
Z′′<br />
xx(x, y) cos 2 a 2Z′′<br />
xy(x, y) sin a cos a Z′′<br />
yy(x, y) sin 2 a<br />
U′′ XY (X,Y) <br />
U′′ YY(X,Y) <br />
Z′′<br />
xx(x, y) sin 2 a 2Z′′<br />
xy(x, y) sin a cos a Z′′<br />
yy(x, y) cos 2 a.<br />
Writing: sin 2A 2Z′′<br />
xy /C, cos 2A (Z′′<br />
xx Z′′<br />
yy )/C, where<br />
C √(Z′′ xx Z′′ yy) 2 2 4Z′′ xy ,<br />
one gets the expressions:<br />
U′′ XX (Z′′ xx Z′′ yy)<br />
<br />
2<br />
C<br />
cos 2(a A)<br />
2<br />
U′′ XY C<br />
sin 2(a A)<br />
2<br />
Z′′<br />
xx (x, y) sin a cos a Z′′<br />
xy (x, y) (cos 2 a sin 2 a)<br />
Z′′<br />
yy (x, y) sin a cos a<br />
U′′ YY (Z′′ xx Z′′ yy )<br />
<br />
2<br />
C<br />
cos 2(a A)<br />
2<br />
(3.30)<br />
(3.31)
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
from which it is easy to verify that<br />
U′′ XX 2 2U′′XY 2 U′′<br />
YY 2 Z′′<br />
xx 2 2Z′′<br />
xy 2 Z′′<br />
yy 2 .<br />
Justification of the approximation of the topographic surface by<br />
a thin plate spline surface<br />
One understands now that the condition of minimality of the integral K(Z)<br />
means that between sample points the ‘global’ curvature of the surface is<br />
as weak as possible, or else that variations of the tangent plane, and therefore<br />
of the slope, are as weak as possible. So a thin plate spline surface<br />
adjusted on contour lines presents a slope as regular as possible between<br />
contour lines, which is compatible with the definition of the topographic<br />
surface.<br />
We should add that the maximal regularity of the thin plate spline<br />
surface applies at a time in all directions, and in particular in the horizontal<br />
direction, which notably has the effect of ‘straightening’ the surface<br />
exaggeratedly on crests and in thalwegs; the adjusted surface may therefore<br />
present artefacts, and then describes the topographic surface only<br />
roughly.<br />
Implementing the approximation by a thin plate spline surface<br />
STEP 1. CONSTITUTION OF AN APPROPRIATE SAMPLE OF POINTS<br />
To use the method, it is necessary to first select on contour lines a sample<br />
of points, while remembering that a segment of curve [p ij , p ii, j ] must<br />
influence the surface by all its points, and not only by its extremities. By<br />
taking as a sample only the initially given points p ij, one sees that all<br />
segments would have the same influence, whereas it seems natural that a<br />
segment influences according to its length. Therefore one should cut all<br />
contours into small elementary ‘segments’ of similar length, in order to<br />
assure an influence proportional to the length.<br />
STEP 2. CHOICE OF WEIGHTS IN THE CASE OF THE THIN PLATE SPLINE<br />
SURFACE OF ADJUSTMENT<br />
A thin plate spline surface of adjustment Z(x, y) by definition minimizes<br />
the quantity:<br />
E(Z) K(Z) i<br />
preferably written<br />
i Z(x i , y i ) z i 2 ,<br />
DSM reconstruction 239
240 Grégoire Maillet et al.<br />
where<br />
r i is the relative weight of the point (x i , y i , z i ) and the P weight fixes the<br />
importance of the adjustment criterion<br />
in relation to the curvature criterion K(Z).<br />
It is then necessary to choose the weight P and weights r i .<br />
(a) Let us first examine the effect of the weight P.<br />
To P > 0, corresponds a unique thin plate spline surface, denoted Z P ,<br />
minimizing E P (Z) K(Z) Pe(Z); in the same way, to Q > 0 corresponds<br />
the surface Z Q , minimizing E Q (Z) K(Z) Qe(Z).<br />
Because of the minimal character of E P (Z P ) and of E Q (Z Q ), one has<br />
Or else:<br />
E P (Z) K(Z) P i<br />
P i<br />
e(Z) i<br />
i and r i i<br />
r i Z(x i, y i) z i 2<br />
E P(Z P) E P(Z Q) and E Q(Z Q) E Q(Z P).<br />
K(Z P) Pe(Z P) K(Z Q) Pe(Z Q),<br />
K(Z Q ) Qe(Z Q ) K(Z P ) Qe(Z P ).<br />
Adding on the one hand the inequalities,<br />
K(Z P) Pe(Z P) K(Z Q) Pe(Z Q).<br />
K(Z P ) Qe(Z P ) K(Z Q ) Qe(Z Q ),<br />
and on the other hand the inequalities,<br />
QK(Z Q) PQe(Z Q) QK(Z P) PQe(Z P),<br />
PK(Z Q) PQe(Z Q) PK(Z P) PQe(Z P),<br />
one gets the inequalities,<br />
P ;<br />
(P Q)e(Z P) (P Q)e(Z Q)<br />
r i Z(x i , y i ) z i 2 ,
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
and<br />
(P Q)K(Z Q ) (P Q)K(Z P ), (3.32)<br />
showing that if P>Q, then K(Z P ) K(Z Q ) and e(Z P ) e(Z Q ); it means<br />
that the surface Z Q is less bent than Z P , or ‘smoother’, but not so well<br />
adjusted to the sample. In particular the interpolation thin plate spline Z ,<br />
limit of Z P when P becomes arbitrarily large, is perfectly adjusted to the<br />
sample, but is less smooth than any thin plate spline of adjustment Z Q.<br />
(b) We suppose the P weight henceforth fixed, and we consider the weights<br />
r i .<br />
The simplest choice is of course to give the same weight r i 1/n to all<br />
points.<br />
This choice is not satisfactory because the distribution of points(x i, y i)<br />
can be heterogeneous because of the irregular horizontal distribution of<br />
contours, and of variations of the sampling interval along contours.<br />
Uniform weights would make the adjustment surface Z(x, y) deflect toward<br />
portions of land finely sampled, which is unacceptable because Z has no<br />
reason to be affected by the sampling interval.<br />
To avoid this inconvenience, it is therefore necessary that weights r i are<br />
small where the sample is dense, and larger where the sample is sparse.<br />
In other words a weight r i must be weak when the point (x i, y i) is representative<br />
of a small area A i, and large when the point is representative of<br />
a large one, which brings about that r i is an increasing function f(|A i|) of<br />
the area |A i|. Besides, this function can only be linear, because if one unifies<br />
two contiguous parcels, A i, A j in one A i ∪ A j, then the weight f (|A i ∪ A j|)<br />
f(|A i| |A j|) of the union must be the sum of weights:<br />
f(|A i | |A j |) r i r j f(|A i |) f(|A j |). (3.33)<br />
Thus, ri k|Ai|, and the condition imposes<br />
ri 1<br />
k <br />
1<br />
A i .<br />
DSM reconstruction 241<br />
It remains to specify what is a parcel A i represented by the point<br />
(x i , y i , z i ); it is natural to define it as the set of points (x, y) that are closer<br />
to (x i, y i) than to the other (x j, y j) points; one knows that this set is a<br />
polygon V i and that the family {V i; i 1, . . ., n} forms the Voronoï diagram<br />
associated with points (x i,y i).<br />
Finally, a coherent system of weight can be defined by:
242 Grégoire Maillet et al.<br />
r i <br />
where |V i | designates the area of the Voronoï polygon V i .<br />
STEP 3. RESOLUTION OF THE LINEAR SYSTEM<br />
(3.34)<br />
The following step consists in solving the linear system of n3 equations<br />
giving coefficients a 1 , . . ., a n , b 0 , b 1 , b 2 .<br />
Denoting c ij r i (x j , y j ) 2 ln r i (x j , y j ) this system becomes in matrix form:<br />
⎧<br />
⎪<br />
⎪<br />
⎪<br />
⎪<br />
⎩<br />
j<br />
8/ 1<br />
c 12<br />
…<br />
c 1n<br />
1<br />
x 1<br />
y 1<br />
which can be written as<br />
C F<br />
T<br />
F 0 a<br />
b Z<br />
0<br />
Note that,<br />
V i<br />
. (3.35)<br />
ri(xj, yj) √(xj xi) ,<br />
2 (yj yi) 2 rj(xi, yi) therefore c ji c ij; the matrix C is therefore symmetrical.<br />
Note that neither the matrix<br />
C F<br />
T<br />
F 0<br />
V j ,<br />
c 21<br />
8/ 2<br />
…<br />
c2n 1<br />
x2 y2 …<br />
…<br />
…<br />
…<br />
…<br />
…<br />
…<br />
nor its opposite, are positive; for example, if a = (1, 0, 0, …, 0) and b =<br />
(b 0 , 0, 0), one gets:<br />
(aT bT ) C F<br />
T<br />
F 0 a<br />
b<br />
c n1<br />
c n2<br />
…<br />
8/ n<br />
1<br />
xn yn 8<br />
2b0 1<br />
1<br />
1<br />
…<br />
1<br />
0<br />
0<br />
0<br />
x 1<br />
x 2<br />
…<br />
x n<br />
0<br />
0<br />
0<br />
which can be 0 or 0 according to the value of b 0 .<br />
C. Carasso in Baranger (1991) suggests solving the system (3.35) as<br />
follows. One factorizes F as F QR where Q is an (n, n) orthogonal<br />
matrix, and R an (n, 3) matrix in the shape<br />
y1 y2 …<br />
y n<br />
0<br />
0<br />
0<br />
⎫<br />
⎪<br />
⎪<br />
⎪<br />
⎪<br />
⎭<br />
⎧<br />
⎪<br />
⎪<br />
⎪<br />
⎪<br />
⎩<br />
a 1<br />
a 2<br />
…<br />
a n<br />
b 0<br />
b 1<br />
b 2<br />
⎫<br />
⎪<br />
⎪<br />
⎪<br />
⎪<br />
⎭<br />
<br />
⎧<br />
⎪<br />
⎪<br />
⎪<br />
⎪<br />
⎩<br />
z1 z2 …<br />
z n<br />
0<br />
0<br />
0<br />
⎫<br />
⎪<br />
⎪<br />
⎪<br />
⎪<br />
⎭
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
R U<br />
0<br />
,<br />
with U a (3, 3) triangular superior matrix. F is of rank 3, since there are<br />
at least three non-aligned points, therefore R is of rank 3, and thus U is<br />
invertible. So, one can write (here I is the (3, 3) identity matrix):<br />
C F<br />
T<br />
F 0 Q<br />
0 0<br />
I QTCQ T<br />
R<br />
R<br />
0<br />
so that the system (3.35) is equivalent to the system<br />
0<br />
(3.36)<br />
(3.37)<br />
Writing by blocks, Q (Q 1 Q 2 ), Q 1 (n, 3) matrix, Q 2 (n, n3) matrix, and<br />
this system is written:<br />
Q1 TCQ1 T<br />
Q2CQ1 U T<br />
T Q1CQ2 T<br />
Q2CQ2 0<br />
0 QT<br />
QTCQ R<br />
T R 0 d<br />
b QTZ 0 , a Qd .<br />
d d 1<br />
d 2 , d 1(3) vector, d 2(n3) vector,<br />
U<br />
0<br />
0 d1 d2 b Q1 TZ (3.38)<br />
one deduces U T d 1 0, thus d 1 0; from the equation Q 2 T CQ2d 2 Q 2 T Z,<br />
one deduces d 2 , from where a Q 2 d 2 ; from the equation Q 1 T CQ2 d 2 Ub<br />
Q 1 T Z, or Ub Q1 T (Z Ca), one finally extracts b.<br />
It is necessary to notice that the C matrix is full, therefore the matrix<br />
Q 2 T CQ2 also; one will see that to the contrary the ‘elastic’ grid method<br />
drives to a sparse system, allowing larger size systems to be processed with<br />
the same memory resources.<br />
STEP 4. CONSTITUTION OF THE REGULAR GRID<br />
T<br />
Q2Z 0 , a Q1d1 Q2d2 ;<br />
This immediate step consists in calculating, by the formula (3.26), the altitudes<br />
z cl Z(x c , y l ) at the grid nodes.<br />
3.5.2.4 Approximation of the topographic surface by an elastic<br />
grid surface (de Masson d’Autume, 1978)<br />
Definition<br />
An elastic grid surface can be introduced as a discrete approximation of<br />
a thin plate spline surface of adjustment Z(x, y), defined in §3.5.2.3; one<br />
knows that Z(x, y) minimizes a quantity:<br />
I<br />
DSM reconstruction 243
244 Grégoire Maillet et al.<br />
E(Z) K(Z) i<br />
(3.39)<br />
As our final objective is not the Z(x, y) surface, but only the set of its<br />
values zcl Z(xc, yl) at the nodes of a regular grid {xc ch; c 1, . . ., N}<br />
× {yl lh; l 1, . . ., M}, one can try to set a problem whose unknown is the<br />
set {zcl } directly. Thus we shall replace the E(Z) quantity by an approached<br />
quantity that is a function of the zcl only.<br />
(1) On the one hand one has to replace the integral<br />
One first approaches it by Riemann sum:<br />
K(Z) h 2 c,l<br />
where c 1:N, l 1:M.<br />
i Z(x i , y i ) z i 2 .<br />
K(Z) Z′′<br />
xx 2 2Z′′<br />
xy 2 Z′′yy 2 dxdy.<br />
Z′′<br />
xx (x c , y l ) 2 2Z′′<br />
xy (x c , y l ) 2 Z′′<br />
yy (x c , y l ) 2 ,<br />
(3.40)<br />
One can then write the following eight Taylor formulae at point (x c, y l):<br />
zc1,l zcl Z′ xh Z′′ xx h2<br />
Z′′′ xxx<br />
2 h3<br />
6 o(h3 )<br />
zc1,l zcl Z′ xh Z′′ xx h2<br />
Z′′′ xxx<br />
2 h3<br />
6 o(h3 )<br />
zc,l1 zcl Z′ yh Z′′ yy h2<br />
Z′′′ yyy<br />
2 h3<br />
6 o(h3 )<br />
zc,l1 zcl Z′ yh Z′′ yy h2<br />
Z′′′ yyy<br />
2 h3<br />
6 o(h3 )<br />
zc1,l1 zcl Z′ xh Z′ yh (Z′′ xx 2Z′′ xy Z′′ yy) h2<br />
2 <br />
h3<br />
(Z′′′ xxx 3Z′′′ xxy 3Z′′′ xyy Z′′′ yyy )<br />
6 o(h3 )<br />
zc1,l1 zcl Z′ xh Z′ yh (Z′′ xx 2Z′′ xy Z′′ yy) h2<br />
2 <br />
(Z′′′ xxx 3Z′′′ xxy 3Z′′′ xyy Z′′′ yyy) h3<br />
6 o(h3 )<br />
zc1,l1 zcl Z′ xh Z′ yh (Z′′ xx 2Z′′ xy Z′′ yy ) h2<br />
2
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
zc1,l1 zcl Z′ xh Z′ yh (Z′′ xx 2Z′′ xy Z′′ yy ) h2<br />
2 <br />
h3<br />
(Z′′′ xxx 3Z′′′ xxy 3Z′′′ xyy Z′′′ yyy ) (3.41)<br />
6<br />
adding the first and second formulae, then the third and fourth, and finally<br />
the last four, one deduces:<br />
o(h3 );<br />
Z′′<br />
xx (x c , y l ) (z c1,l 2z cl z c1,l)<br />
h 2 o(h)<br />
Z′′<br />
yy (x c , y l ) (z c,l1 2z cl z c,l1)<br />
h 2 o(h)<br />
(3.42)<br />
These values permit us to approximate the integral K(Z) by the quantity<br />
function of the vector z {zcl} of RMN Z′′ xy (xc , yl ) <br />
:<br />
(zc1,l1 zc1,l1 zc1,l1 zc1,l1) 4h 2 o(h).<br />
K h(z) <br />
1<br />
h 2 N1<br />
c2<br />
N1<br />
c2<br />
(2) On the other hand one has to replace in<br />
e(Z) i<br />
M1<br />
(zc1,l1 zc1,l1 zc1,l1 zc1,l1) l2<br />
2 /8 .<br />
(3.43)<br />
every Z(x i , y i ) by a function of the vector z.<br />
One will adopt a representation by a piecewise bilinear or bicubic function<br />
Z(x, y) c,l<br />
M<br />
(zc1,j 2zcl zc1,l) l1<br />
2 N<br />
c1<br />
i Z(x i, y i) z i 2<br />
zcl V x<br />
h c V y<br />
l h<br />
with, in the bilinear case, V(t) Q(t) sup(1 |t |, 0) and in the bicubic<br />
case:<br />
3<br />
2 t<br />
V(t) U(t) 3 5<br />
2 t 2 1<br />
1<br />
2 t 3 5<br />
2 t 2 4t 2<br />
0<br />
DSM reconstruction 245<br />
(Z′′′ xxx 3Z′′′ xxy 3Z′′′ xyy Z′′′ yyy) h3<br />
6 o(h3 )<br />
M1<br />
(zc,l1 2zcl zc,l1) l2<br />
2<br />
if t 1<br />
if 1 t 2<br />
if t 2<br />
(3.44)
246 Grégoire Maillet et al.<br />
e(Z) is then replaced by<br />
e h(Z) i<br />
Finally the E(Z) quantity is replaced by the quantity function of<br />
One defines the elastic grid surface associated with the sample {(x i, y i, z i);<br />
i 1, . . ., n} as the surface<br />
Z(x, y) c,l<br />
defined by the vector z {z cl } that minimizes the quantity E h (z).<br />
Existence and unicity of the solution for the minimization<br />
problem of Eh(z) The quantity Eh(z) is a quadratic function of z {zcl}, which therefore can<br />
be minimized explicitly.<br />
In a precise way, one introduces the vectors Ccl ,Dcl ,Fcl ,Bi such as<br />
c,l<br />
for example,<br />
C T cl<br />
= (0 … 0 1 –2 1 0 … 0) .<br />
One then sees that:<br />
e h(z) i<br />
i c,l<br />
z {z cl}, E h(z) K h(z) e h(z).<br />
T zc1,l 2zcl zc1,l Ccl z<br />
T zc,l1 2zcl zc,l1 Dcl z<br />
K h(z) 1<br />
h 2 cl<br />
z T i<br />
z cl V x i<br />
h c V y i<br />
h l z i 2<br />
.<br />
zcl V x<br />
h c V y<br />
l h<br />
z c1,l1 z c1,l1 z c1,l1 z c1,l1 F cl T z<br />
zcl V xi h c V yi h l BT i z ,<br />
c1,l cl c1,l<br />
zT T CclCcl z cl i (z T B i z i) (B i T z zi) <br />
i B i B i T z 2 i<br />
zT T 1<br />
DclDcl z <br />
8 cl i z i B i T<br />
z i<br />
zT T FclFcl z<br />
i z i 2 ,<br />
(3.45)<br />
(3.46)
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
so that, putting:<br />
A 1<br />
h 2 cl<br />
b i<br />
one can write E h (z) as:<br />
The matrix A is symmetrical. It is positive because<br />
(3.47)<br />
whatever the value of z.<br />
One intends to show that the matrix A is positive definite, which means<br />
that z T Az can be 0 only if z 0; this property will guarantee the existence<br />
and the unicity of the solution of the problem.<br />
Let us suppose therefore that z T Az for a certain z.<br />
Then C T cl z 0 for any c, l, which implies z cl z 1l (c 1)(z 2l z 1l)<br />
In the same way D T cl z 0 for any c, l, which implies z cl z c1 (l 1)<br />
(z c2 z c1).<br />
One deduces then:<br />
zcl z11 (c1)(z21 z11) (l1)(z12 z11) <br />
(c1)(l1)(z22 z12 z21 z11 )<br />
a0 a1c a2l a3cl (3.48)<br />
for any c, l;<br />
from where<br />
T CclCcl cl i z i B i c i<br />
E h (z) z T Az 2b T z c.<br />
zTAz 1<br />
h2 T 2 (Ccl z) cl<br />
cl<br />
i<br />
i (B i T z) 2 0<br />
T 1<br />
DclDcl <br />
8 cl i z i 2<br />
T 2 1<br />
(Dcl z) <br />
8 cl F T<br />
cl z z c1,l1 z c1,l1 z c1,l1 z c1,l1 4a 3 .<br />
The hypothesis z T Az 0 also implies F T cl z for any c, l, from where a 3 0.<br />
Finally: z cl a 0 a 1c a 2l for any c, l.<br />
DSM reconstruction 247<br />
F T clFcl i T (Fcl z) 2<br />
iB iB i T
248 Grégoire Maillet et al.<br />
One has then:<br />
B i T z<br />
<br />
However, one can show that the V function, when it designates the Q<br />
linear interpolation function as well as the U cubic interpolation function,<br />
verifies the identities:<br />
c<br />
cl<br />
from where:<br />
a 0 c<br />
But the hypothesis z T Az 0 implies B i T z a0 a 1(x i/h) a 2(y i/h) 0<br />
for any i; now one can suppose that the sample counts at least three (x j,<br />
y j), (x k, y k), (x p, y p) non-aligned points; these three points verify the system<br />
a xj 0 a1 a 1 c<br />
a 2 c<br />
V(x c) 1 and c<br />
B i T z a0 a 1 x i<br />
h a 2 y i<br />
h .<br />
y j<br />
h a2 0,<br />
h<br />
cV x i<br />
h c l<br />
V x i<br />
h c l<br />
of non-null determinant (x p – x j )(y k – y j ) – (x k – x j )(y p – y j ), which is<br />
possible only if a 0 a 1 a 2 0. Therefore z 0; one has thus shown<br />
that A is positive definite.<br />
A is then invertible, and one can write therefore:<br />
E h (z) (z A 1 b) T A(z A 1 b) c b T A 1 b c b T A 1 b,<br />
which shows that E h(z) reaches its minimum cb T A 1 b if and only if z <br />
A 1 b. In other words the problem of minimization of E h(z) admits a solution<br />
z and only one given by the equation Az b.<br />
Structure of the A matrix<br />
One has: A A1 A2 where<br />
(a0 a1c a2l) V xi h c V yi l h<br />
V x i<br />
h c l<br />
V yi l h<br />
V yi l h<br />
lV yi l .<br />
h<br />
cV(x c) x for any x;<br />
xk a0 a1 h a yk 2 0,<br />
h<br />
xp a0 a1 h a yp 2 0 h
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
A 1 1<br />
h 2 cl<br />
A, A 1, A 2 are (MN, MN) matrixes (M: number of lines, N: number of<br />
columns of the grid).<br />
If we expand A, we get:<br />
A 1 1<br />
h 2<br />
where I is the (N, N) identity matrix, X and Y are (N, N) matrixes:<br />
X <br />
Y <br />
Expanding A 2 in the case of bilinear model one gets:<br />
A 2 <br />
⎧<br />
⎪<br />
⎪<br />
⎪<br />
⎪<br />
⎩<br />
⎧<br />
⎪<br />
⎪<br />
⎪<br />
⎪<br />
⎩<br />
⎧<br />
⎪<br />
⎩<br />
1<br />
2<br />
1<br />
1<br />
0<br />
1<br />
⎧<br />
⎪<br />
⎪<br />
⎩<br />
IX Y<br />
8<br />
X<br />
X<br />
2I<br />
I Y<br />
8<br />
2<br />
5<br />
4<br />
…<br />
0<br />
1<br />
0<br />
…<br />
X<br />
X<br />
…<br />
C cl C T cl cl<br />
1<br />
4<br />
6<br />
…<br />
1<br />
1<br />
0<br />
2<br />
…<br />
1<br />
X<br />
…<br />
X<br />
2I<br />
5IX Y<br />
8<br />
4I<br />
…<br />
1<br />
0<br />
…<br />
0<br />
1<br />
…<br />
X<br />
X<br />
1<br />
4<br />
…<br />
4<br />
1<br />
1<br />
…<br />
2<br />
0<br />
1<br />
X<br />
X<br />
D cl D T cl 1<br />
8 cl<br />
1<br />
…<br />
6<br />
4<br />
1<br />
⎫<br />
⎪<br />
⎪<br />
⎭<br />
I Y<br />
8<br />
4I<br />
6IX Y<br />
4<br />
…<br />
I Y<br />
8<br />
…<br />
4<br />
5<br />
2<br />
…<br />
0<br />
1<br />
0<br />
1<br />
0<br />
1<br />
I Y<br />
8<br />
4I<br />
…<br />
…<br />
I Y<br />
8<br />
1<br />
2<br />
1<br />
⎫<br />
⎪<br />
⎪<br />
⎪<br />
⎪<br />
⎭<br />
⎫<br />
⎪<br />
⎪ .<br />
⎪<br />
⎪<br />
⎭<br />
DSM reconstruction 249<br />
F cl F T cl ,<br />
I Y<br />
8<br />
…<br />
6IX Y<br />
4<br />
4I<br />
I Y<br />
8<br />
A 2 i<br />
…<br />
4I<br />
5IX Y<br />
8<br />
2I<br />
i B i B i T<br />
…<br />
I Y<br />
8<br />
2I<br />
IX Y<br />
8<br />
,<br />
⎫<br />
⎪<br />
⎭
250 Grégoire Maillet et al.<br />
where each X is a matrix of the shape<br />
X <br />
⎧<br />
⎪<br />
⎪<br />
⎩<br />
•<br />
•<br />
A 1 and A 2 are written in the shape of (M, M) matrixes of (N, N) size<br />
blocks.<br />
Every line of the matrix A includes a maximum of 21 non-null coefficients<br />
on a total of MN; for a grid M 100, N 100, there are therefore<br />
more than 99.79 per cent of zero coefficients. One sees that A is really a<br />
sparse matrix, as stated.<br />
References<br />
•<br />
•<br />
…<br />
•<br />
…<br />
•<br />
…<br />
•<br />
•<br />
•<br />
•<br />
⎫<br />
⎪<br />
⎪<br />
⎭<br />
Carasso C. (1991) Lissage des données à l’aide de fonctions spline, in J. Baranger<br />
(ed.) Analyse numérique, Hermann, pp. 357–414.<br />
de Masson d’Autume G. (1978) Construction du modèle numérique d’une surface<br />
par approximations successives. Application aux modèles numériques de terrain<br />
(M.N.T.), Bulletin de la Société française de photogrammétrie et de télédétection,<br />
no. 71–72, pp. 33–41.<br />
de Masson d’Autume G. (1979) Surface modelling by means of an elastic grid,<br />
Photogrammetria, no. 35, pp. 65–74.<br />
Duchon J. (1976) Interpolation des fonctions de deux variables suivant le principe<br />
de la flexion des plaques minces, Revue Française d’Automatique, Informatique<br />
et Recherche Opérationnelle (R.A.I.R.O.) Analyse numérique, vol. 10, no. 12,<br />
décembre, pp. 5–12.<br />
3.5.3 Merging contour lines and 3D measurements from<br />
images<br />
Nicolas Paparoditis<br />
The two techniques that have been described in the previous paragraphs,<br />
for understandable pedagogic and clarity reasons, have been developed<br />
separately. Nevertheless, the DSM reconstruction from contour lines can<br />
be improved by adding 3D points processed by photogrammetric means to<br />
solve ambiguities between all possible triangulation of contour lines. In the<br />
same way, contour lines can be injected as constraint lines in the process<br />
of triangulation of all 3D points that can processed from the images. Both<br />
techniques lead to a constrained triangulation. Nevertheless, these merging<br />
processes will have the chance of succeeding if both 3D information are<br />
homogeneous in accuracy and precision and definitely describe the same<br />
surface.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3.5.4 Generating a DTM with sparse 3D points and<br />
characteristic features<br />
Nicolas Paparoditis<br />
DSM reconstruction 251<br />
A DTM can be generated manually by plotting well-chosen sparse on<br />
ground 3D points in a stereo display. The spatial localization of these<br />
points is determined by the operator and their elevation can be also manually<br />
plotted or automatically generated through the stereo matching<br />
methods we have discussed earlier. The spatial density of the points depends<br />
on the local roughness of the landscape and on the desired DTM quality.<br />
A TIN-like surface can be generated from these points with a 2D Delaunay<br />
triangulation, as shown in Figure 3.5.4.1 (a) and (b).<br />
Instead of plotting much denser points (increasing production times)<br />
around terrain break lines to render locally the relief morphology, the<br />
operator can plot the break lines themselves and use these lines to constrain<br />
locally the triangulation. In practice, we destroy all triangles overlapping<br />
x<br />
y<br />
(a) (b)<br />
f g<br />
j<br />
k a<br />
i<br />
b h<br />
l<br />
c<br />
m<br />
n<br />
d a<br />
e<br />
o<br />
p<br />
(c) (d)<br />
Figure 3.5.4.1 (a) in black sparse cloud of 3D points and in grey the slope beak<br />
line; (b) 2D triangulation of all points in the cloud; (c) all triangles<br />
overlapping the break line are removed and (d) the remaining<br />
space is triangulated under the constraint that the new triangles lie<br />
on the break line segments.
Figure 3.5.4.2 (a) Triangulation without characteristic constraint lines.<br />
Figure 3.5.4.2 (b) Triangulation with characteristic constraint lines.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Extraction of characteristic lines of the relief 253<br />
the break line and we triangulate the remaining space. For instance, in the<br />
case of an open slope break line as shown on Figure 3.5.4.1, the triangulation<br />
of this empty space can be seen as the triangulation of the polygon<br />
(c, f, g, h, k, l, m, n, o, p, f, c, d, e, d, c, b, a, b, c) in the Figure<br />
3.5.4.1 (c) example, where c f is the smallest segment joining the break<br />
line points and the surrounding triangle points.<br />
Figures 3.5.4.2 (a) and (b) show some triangulation results on a real set<br />
of data (points and characteristic terrain line) plotted on a digital photogrammetric<br />
workstation.<br />
3.6 EXTRACTION OF CHARACTERISTIC LINES OF THE<br />
RELIEF<br />
Alain Dupéret, Olivier Jamet<br />
3.6.1 Some definitions bound to the relief<br />
3.6.1.1 Preliminary definitions<br />
The building of the relief results from many endogenous actions and exogenous<br />
phenomena such as lithology, climatic variations, vegetation,<br />
anthropic activities. . . . The hydrographic network is a set of particularly<br />
useful morphological objects to characterize the relief. The description of<br />
this last by the characteristic elements leans on a mathematical representation<br />
that depends on:<br />
• the type of data capture: the surface itself, the cartographic representation;<br />
• the mode of acquisition: direct surveys, digitization, photogrammetric<br />
restitution;<br />
• the mode of altitudes distribution: contour line, regular or irregular<br />
set of points.<br />
In practice, it is desirable to do a separation between aspects bound to<br />
the acquisition of the ground characteristic elements and those bound<br />
to the mathematical representation of the surface. Otherwise, notions of<br />
thalwegs and rivers are often assimilated, and this leads to confusions<br />
such as how to assimilate the relief and the process of water streaming,<br />
even if they are the most often linked; the superficial outflow most often<br />
takes place in the sense of the local or regional line of stronger slope, but<br />
the geology (by the presence of hard or soft rock) and the fracturation<br />
can also intervene.<br />
First, a simple topological mathematical presentation will be given for<br />
the main remarkable elements of the relief. Then, several operative modes<br />
will be presented on the basis of some important studies.
254 Alain Dupéret and Olivier Jamet<br />
3.6.1.2 Definitions of the main characteristic elements of the<br />
relief<br />
The definitions that follow are based upon a quite simplistic modelization<br />
of the reality: we will suppose that the topographic surface is correctly<br />
described by a function z H(x, y), continuous and twice derivable.<br />
Obviously it is only an approximation, as the topography presents many<br />
similarities with fractal structures, and presents discontinuities at every<br />
scale of observation. Nevertheless, this approach allows for much simpler<br />
definitions of the characteristic elements of the relief.<br />
The main remarkable points<br />
THE SUMMIT<br />
The summit is the high point isolated representative of the local maximum<br />
of the H function. In practice, one generally uses the following (incomplete)<br />
characterization:<br />
THE BASIN<br />
The basin is the low point isolated representing the local minimum of the<br />
H function. In practice:<br />
which means therefore that radii of curvatures in the principle planes are<br />
negative.<br />
THE PASS<br />
(X 0 , Y 0 ) is a summit if H′ x(X 0, Y 0) 0<br />
H′ y (X 0 , Y 0 ) 0<br />
H′′<br />
xx(X 0, Y 0) < 0 and H′′<br />
yy (X 0, Y 0) < 0<br />
(X0 , Y0 ) is a basin if H′ x(X0, Y0) 0<br />
H′ y (X0 , Y0 ) 0<br />
,<br />
H′′ xx(X0, Y0) > 0 and H′′ yy (X0, Y0) > 0<br />
The pass is an origin point of at least two divergent water streams. The<br />
notion of pass evokes a point as much as a surface. From the macroscopic<br />
point of view, it is a region in which two divergent valleys separating two<br />
summits meet that is, mathematically, a saddle point. In practice:<br />
(X0, Y0) is a pass if H′ x (X0 , Y0 ) 0<br />
H′ y (X0, Y0) 0<br />
H′′<br />
xx(X 0, Y 0) · H′′<br />
yy (X 0, Y 0) < 0<br />
.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
This situation implies the existence of such a point, that any infinitesimal<br />
circle around it shows at least a double slope alternance. The water stream<br />
may, from a minute displacement, go in two different directions.<br />
The detection of passes according to these criteria often produces in<br />
practice many parasitic passes, of weak amplitude, close to an extremum<br />
of H. Most often, a minimal modification of altitudes in the DTM makes<br />
such artefacts disappear. The replacement of a square mesh by a triangular<br />
mesh would also annul a certain number of them. This purification of<br />
the model is a prerequisite for a correct topographic modelling and an<br />
easier extraction of the characteristic lines.<br />
The main characteristic lines<br />
From a global point of view, a characteristic line is a line of slope assuring<br />
the sharing of the outflow of the streaming waters. The definitions of these<br />
lines often present two variants, depending on whether they are understood<br />
according to the real geomorphology, or according to the algorithms<br />
used to detect them. For example, with the geomorphological meaning,<br />
a thalweg is generally not formed at the beginning of the watershed of a<br />
torrent, the opposite of what the determination algorithms may show, as<br />
they may define a ‘theoretical’ thalweg that may climb up to the pass.<br />
THE MAIN THALWEG<br />
The thalweg can be determined by the journey of a stream between the<br />
pass that is maybe at its origin, to its end naturally in the basin that it is<br />
going to find on its journey, while following the line of stronger downward<br />
slope.<br />
THE MAIN CREST<br />
Extraction of characteristic lines of the relief 255<br />
The localization of every crest, in the same way, is achieved as being the<br />
journey followed by an anti-streaming (i.e. reversed streaming, assimilated<br />
to water that would go up exclusively along the line of stronger slope) from<br />
a pass to a summit, while following the line of stronger ascending slope.<br />
CONCLUSION ON THE MAIN CHARACTERISTIC LINES<br />
Crests and main thalwegs constitute the main oro-hydrographic system<br />
that relies completely on the preliminary determination of the passes.<br />
Special attention must be paid to low and/or flooded zones such as marshes,<br />
lakes and large flat valleys, in particular when automatic extraction is<br />
concerned.<br />
All goes well if a tracing of a crest started in a convex zone is constituted<br />
in a convex zone up to the junction with another crest or up to a summit,
256 Alain Dupéret and Olivier Jamet<br />
but problems arise if there is an extension in a concave zone. A symmetric<br />
remark can be made for a thalweg. This justifies the definition and the<br />
taking into account of new characteristic elements of the relief.<br />
3.6.1.3 Definitions of remarkable elements of the relief<br />
The contour line<br />
Contour lines are the plane curves of the function Z H(x, y). Associated<br />
to the selected and regularly spaced altitudes of an interval called equidistance<br />
of curves, they are often used to provide the cartographic<br />
representation of the relief, often with a set of edition spot heights distributed<br />
on remarkable planimetric points. Beyond this cartographic setting,<br />
the horizontals can become characteristic elements of the relief as soon as<br />
it contains extents of water. Contours of ponds and the foreshores of seas<br />
are typically lines that present a remarkable slope discontinuity in the altitude<br />
function for large topographic areas.<br />
The curve n 0<br />
The horizontal curvature of contour lines can also become the object of<br />
an analysis. The convention, of course, foresees that an observer who<br />
constantly follows a contour line in the direct sense turns on his right in<br />
convex ground (left in concave). Numerous difficulties, notably those bound<br />
to the parasitic passes, can be avoided by considering the limit between<br />
the concave and convex zones of the ground. While calling n the horizontal<br />
curvature of contour lines, the curvature n 0 represents the limit<br />
we look for. In principle, the pass is a part of this particular curve.<br />
The secondary remarkable points<br />
Points whose definitions follow correspond to places of null curvature.<br />
THE THALWEG ORIGIN<br />
To palliate annoyances mentioned previously concerning the main characteristic<br />
lines, the high points of the curve n 0 will be taken as thalweg<br />
origins.<br />
THE END OF CREST<br />
The crest ends will be defined as the low points of the curve n 0.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
REMARKS ON THE SECONDARY REMARKABLE POINTS<br />
Other approaches than the one presented are possible, as the one that<br />
would consist in taking all points for which the slope p is maximum; ends<br />
of crest are indicated when n > 0, beginnings of thalweg by n 0, from the<br />
secondary crest ends, while going up the long of the larger slope line up<br />
to the junction with a main line.<br />
REMARKS ON THE SECONDARY CHARACTERISTIC LINES<br />
The reliable extraction of the characteristic elements has therefore a particular<br />
importance, whose automation appears desirable, especially because<br />
the manual extraction seems delicate. Indeed, in opposition to the acquisition<br />
of level contour lines for which the z device is blocked at the time<br />
of the acquisition, or to the seizure of nodes of a net for which the planimetric<br />
position is imposed, the direct photogrammetric acquisition does<br />
not show any systematism. The positioning freedom is large, extremely<br />
correlated to the altimetric precision of pointing and intervenes in zones<br />
that require a significant effort to be correctly described: vegetation in<br />
thalweg bottoms, flat crests, etc.<br />
3.6.1.4 Notions of drainage network and watershed<br />
A drainage network is defined (Deffontaines and Chorowicz, 1991) as being<br />
composed of ‘the set of the topographic surfaces situated below all neighbouring<br />
high points, generally flowing out according to a direction. These<br />
surfaces can contain water in temporary or permanent manner’. It includes:<br />
• thalwegs: valley bottoms, narrow or large, with water or dry;<br />
• closed endorheic or exorheic depressions such as marshes, sink-holes,<br />
lakes.
258 Alain Dupéret and Olivier Jamet<br />
To do some automatic extractions following this definition, several families<br />
of methods are used on DTM, either by dynamic analysis of the streaming,<br />
or by local analysis of the surface curvature.<br />
In hydrology, the true catchment area relative to a point is defined as<br />
the totality of the topographic surface drained by this point; it is therefore<br />
a geographical entity on which the ensemble of water enters, due to<br />
the rain, and shares and organizes itself to lead to a unique outlet in the<br />
exit of a catchment area. As for the topographic catchment area relative<br />
to a point, this is the location of the points of the topographic surface<br />
whose line of larger slope passes by this point. There is not necessarily<br />
coincidence of the two types of catchment areas, for example for karstic<br />
reliefs where water that streams inside a catchment area can be found in<br />
the exit in another catchment area because of losses and re-emergences.<br />
In the same way, water that infiltrates the ground can meet the impervious<br />
layers that direct it to water-tables participating thus in other<br />
hydrologic basin balance. To finish, human activity generates obstacles or<br />
reaches that often modify the natural logic of the stream.<br />
As the topographic, geological and pedological properties of a catchment<br />
area constitute some essential parameters for its survey, it is nesessary<br />
to perform a reliable digital modelling of the relief, in particular of the<br />
drainage network. In such a context, many digital indexes, more thematic,<br />
are used to characterize a catchment area: the drained surface, lengths of<br />
drainage, the perimeter, the distance to the outlet, the size of the drained<br />
surfaces, densities and frequency of drainage, the indication of compactness,<br />
the ratio of confluence or concentration, the Beven index, etc.<br />
[Depraetère and Moniod, 1991].<br />
3.6.2 Extraction of thalwegs from starting points<br />
This method, named ‘structuralist approach’ by the author (Riazanoff<br />
et al., 1992) is inspired by the physical model of water streaming on a<br />
relief and proposes a dynamic method of determination of crest lines. The<br />
tracing of lines is judiciously dynamic from the starting points chosen,<br />
while following the line of larger slope up to arrive either in a border of<br />
the zone, a local minimum or on an already drawn line. The algorithm<br />
proceeds by two distinct steps. The first consists in describing completely<br />
the network coming down from all main passes and following the largest<br />
slope. The second applies to correcting shortcomings produced by depressions<br />
while forcing the drainage toward the lowest pass.<br />
The thalweg is defined as ‘concave place of water convergence’. The<br />
crest is defined in a dual manner as ‘convex place of anti-streaming convergence’.<br />
These two characteristic line types cross themselves in singular<br />
points, mainly passes, but also the local extrema, the high points (local<br />
maximum of a concave zone) and the low points (minimum local of a<br />
convex zone).
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Extraction of characteristic lines of the relief 259<br />
A constraint of initial progession is applied for the particular case of<br />
the starting point. Three different advance constraints have been used for<br />
this method, the description being made for the extraction of crests.<br />
1 The algorithm of the stream: the constraint of advance is ‘to climb<br />
according to the line of larger slope’, while choosing among neighbours<br />
the one that presents the largest denivelation in relation to the<br />
displacement among the neighbouring points. The starting point is a<br />
pass. Thus the network obtained can be considered as too dense. The<br />
detection is also meaningful for zones situated at a low altitude, if<br />
compared to zones of average or high altitude.<br />
2 The algorithm of the wanderer: from the local maxima, the constraint<br />
of advance used here is ‘to go down towards points presenting a convex<br />
slope change in one of the three directions (four minus the one of<br />
origin)’. This process makes it possible to survey the main structures<br />
but is very sensitive to the least disconnection that leads to the nonconsideration<br />
of possible structure situated downstream.<br />
3 The main pass algorithm adds up the information from the two<br />
previous algorithms in order to select starting points susceptible of<br />
generating the main characteristic lines. From the passes not marked<br />
in the network obtained by the previous method, the constraint of<br />
advance is ‘to climb according to the line of larger slope’. All marked<br />
passes are passes not belonging to the network generated by the<br />
climbing from the local minima on thalwegs; in fact, a pass will not<br />
be marked if it is located on the extremity of a branch of the network.<br />
The main crests are correctly marked. The principal massifs and hills<br />
are marked.<br />
The third method, using the first two, gives best results. The outlet is<br />
searched for. Logically, the point of outlet is the lowest pass among those<br />
from where come lines of outflow arriving at this depression. To improve<br />
the method, the sense of outflow of the line from the outlet is reversed,<br />
and the outflow continues on the other side of the pass. Other approaches<br />
preserved the principle to apply a method of streaming according to the<br />
line of larger slope, without trying to start again from particular points<br />
such as passes. From every pixel, an outflow on the model is simulated<br />
according to the nearest pixel of the line of larger slope.<br />
3.6.3 Extraction of thalwegs by global outflow calculation<br />
or accumulation<br />
Methods that follow exploit the same operative definition of thalwegs as<br />
techniques of progress previously presented: thalwegs, generated by<br />
phenomena of natural erosion of the ground, are places of river passage,<br />
assimilated to lines, understood this time as places where the pluvial waters
260 Alain Dupéret and Olivier Jamet<br />
concentrate. This definition is no longer in order to perform a local extraction<br />
of the network (either by local operators, or by following step by<br />
step a given line), but to exploit the whole surface to determine lines on<br />
which the strongest debits would appear in case of a uniform rain on the<br />
site.<br />
The representation by a regular rectangular mesh is the one that leads<br />
to the simplest algorithms and will be the only one treated in this section<br />
(explanation data are even more strictly restricted to the square meshes).<br />
We will even speak of pixels (by reference to the terminology of Image<br />
Processing) to designate the points of the mesh.<br />
In spite of these limitations, the methods present here the two interests<br />
of offering continuous line extraction and permitting some extremely simple<br />
implementations. As the results are very realistic on all land with a marked<br />
relief, these methods are among the most used.<br />
3.6.3.1 Calculation of the drained surfaces<br />
If we suppose there is a uniform rain across the whole of the land, the debit<br />
of the permanent regime is in any point proportional to the cumulative<br />
surface upstream of this point. Used in its discrete approximation, this<br />
property helps us calculate, for every pixel, its drained surface as the sum<br />
of the surface of a mesh (influence of the pixel considered) and the drained<br />
surfaces of pixels neighbouring, uphill, the considered pixel. Thalwegs are<br />
then local extrema places, on curves of constant altitude, of the drained<br />
surfaces. (See Figure 3.6.1.)<br />
This definition gives rise to several algorithms, depending on whether<br />
one defines neighbourhoods in 4 or in 8-connexity, according to the way<br />
the relation ‘being uphill of’ is defined, a relation that we call relation of<br />
outflow, and according to the technique used to extract lines of thalwegs<br />
from the drained surface image.<br />
We will restrict the explanation to the 4-connexity topology. The use<br />
of the 8-connexity, sometimes recommended, causes the problem of a incoherent<br />
topology: paths of outflows materialized locally by relations of<br />
neighbourhoods cross themselves, which doesn’t correspond to a coherent<br />
physical model. Let us note that the techniques presented below are easily<br />
transposable in 8-connexity, maybe to the cost of an a posteriori correction<br />
procedure of resulting artefacts, which are always local.<br />
The choice of the outflow relation leads to the distinction between two<br />
families of algorithms, presented briefly here. The techniques of extraction<br />
of thalwegs lines themselves, that can be chosen independently of the algorithm<br />
used for the calculation of surfaces, are presented in a different<br />
section.<br />
The simplest approach consists in considering that every pixel cannot<br />
be uphill of more than one pixel among its neighbours. The downstream<br />
pixel will be chosen, on the one hand, so that its altitude is strictly lower
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 3.6.1 Catchment area of a point.<br />
Extraction of characteristic lines of the relief 261<br />
than that of the pixel considered, and on the other by a criterion of slope<br />
(the outflow of waters following the steepest slope).<br />
This formulation ensures that the graph of the outflow relation is without<br />
closed cycle, in other words that this relation defines a tree on the DTM.<br />
(See Figure 3.6.2.)<br />
The calculation of drained surfaces can be treated with a recursive algorithm<br />
or, more elegantly, using the natural order of altitudes to start the<br />
calculation with the highest points, and accumulating drained surfaces in<br />
one pass.<br />
In the formulation of Le Men (Le Roux, 1992), the downstream pixel<br />
is the pixel for which the slope is steepest. This choice is operative on<br />
hilly landscapes (strong local variations of slopes), but produces drifts<br />
on smooth surfaces, bound to the quantification of outflow directions: the<br />
algorithm presents a defect of isotropy, a function of the direction of<br />
the mesh.
262 Alain Dupéret and Olivier Jamet<br />
Figure 3.6.2 Tree of the outflow relation on a DTM.<br />
Contour lines<br />
Pixel considered as an<br />
area<br />
Flow link<br />
(from up toward<br />
downstream<br />
The method of Fairfield (Fairfield and Leymarie, 1991) palliates this<br />
defect by proposing a random choice of the pixel downstream between<br />
the two pixels materializing the two larger slopes, the probability of choice<br />
of one of the two points being defined as proportional to the slope. This<br />
rule guarantees the identity of the mean direction of the pixel downstream<br />
and of the true direction of the steepest slope on a tilted plane, whatever<br />
is the orientation of the mesh. (See Figure 3.6.3.)<br />
Calculation on the graph of outflow<br />
The anisotropy (or the sensitivity to the direction of the mesh) resulting from<br />
the unique choice of the downstream pixel the steepest slope is the origin<br />
of a second family of approaches. Considering that the average direction of
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Contour lines<br />
Pixel considered as an<br />
area<br />
Flow link<br />
(from up toward<br />
downstream)<br />
Extraction of characteristic lines of the relief 263<br />
Figure 3.6.3 Outflow relation on a planar surface: (a) highest slope choice: the<br />
flow drifts away from the real direction; (b) two higher slopes<br />
random choice (Fairfield method): the average flow follows the real<br />
slope.<br />
the outflow must remain faithful to the real direction of the steepest slope,<br />
Rieger (1992) proposed to take into account, as pixels downstream of a<br />
given pixel, all neighbours of lower altitude, and to propagate the surface<br />
drained of a pixel to its downstream pixels according to the slope values.<br />
In this formulation, the relation of outflow defines on the DTM a graph<br />
that is not other than the graph of neighbourhood oriented in the direction<br />
of decreasing altitudes (with exceptions of altitude equality, which we will<br />
return to). This oriented graph being without cycle, as the tree of outflow<br />
discussed previously, the algorithm of drained surface calculation can proceed<br />
in the same way (propagation of drained surfaces in the direction of<br />
decreasing altitudes).<br />
This empiric formulation, which has the advantage of producing a more<br />
regular image of drained surfaces than Fairfield’s method, does not,<br />
however, lead to an isotropic algorithm. A variant in 8-connexity due to<br />
Freeman (1991), founded on an ad hoc formulation of transmitted surfaces,<br />
leads to a better isotropy on surfaces of revolution, but not in the general<br />
case.<br />
Jamet (1994) showed that, in the formulation of Rieger, the calculated<br />
drained surface is considered to be the calculated debit in the permanent<br />
regime, under a uniform rain on the whole DTM, and for a modelling in<br />
which pixels are assimilated to cells communicating by their lateral sides.<br />
For such a modelling, this debit depends on the orientation of the mesh<br />
in relation to the real slope of the ground. The simple calculation of the<br />
debit by unit of length orthogonal to the slope, deduced from the debit<br />
of the mesh and slopes to its neighbours, insures the exact isotropy of the<br />
result. Besides, on regular surfaces, this calculation of isotropic debit<br />
permits a rigorous definition of the convexity of the surface, the gradient<br />
of this debit being equal to the unit on the plane, lower on concave and<br />
greater on convex surfaces. (See Figure 3.6.4.)
264 Alain Dupéret and Olivier Jamet<br />
Figure 3.6.4 Drained surfaces computed on a spherical cap: (a) highest slope<br />
flow; (b) Fairfield method; (c) Rieger method; (d) Freeman method;<br />
(e) Jamet method; (f) theoretical expected result; (g) shows slight<br />
anisotropic effects on non-circular symmetric surfaces with Freeman<br />
method; (h) shows on the same surface as (g) the results of Jamet<br />
method.<br />
Extraction of thalwegs<br />
In all previous methods, the result of the phase of surface accumulation<br />
(or of calculation of the debit) leads to a map superposable on the DTM<br />
on which the researched thalwegs correspond to local extrema orthogonally<br />
to the calculated slope on the DTM.<br />
Three types of techniques are used. The first relies on the classic techniques<br />
of image processing. Considering that the image of surfaces drained<br />
(or of the debit) is contrasted enough (or can become so after application<br />
of a digital skeletonization process), the binary mesh map of thalwegs is<br />
obtained by simple threshold of surfaces. The threshold corresponds<br />
to the minimal size of catchment area able to feed a drain. A vector representation<br />
of the network is deduced by a classic technique of vectorization<br />
(Le Roux, 1992). This method remains more suitable to approaches by<br />
outflow tree, as far as only these guarantee that the surface remains strictly<br />
increasing from the upper to the lower part of a drain (and therefore that<br />
the threshold won’t be able to break their connectivity).<br />
The second uses the tree of outflow as it has been defined previously,<br />
and remains therefore also more adapted to the corresponding family of<br />
methods. The tree of outflow is a vector structure linking the set of pixels<br />
of the DTM. The extraction of the thalweg network consists therefore
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Extraction of characteristic lines of the relief 265<br />
Figure 3.6.5 Drained surfaces computed on a DTM: (a) DTM (shaded);<br />
(b) highest slope algorithm; (c) Jamet’s algorithm.<br />
just in selecting a sub-tree of the outflow tree (Le Roux, 1992). The criterion<br />
of selection will be in the same way as previously a simple threshold of<br />
the drained surface (with the same guarantee of topological consistency).<br />
The third constructs the vector representation of thalwegs in a recursive<br />
way, while starting with stronger debit drains. Except in an exceptional<br />
case, any thalweg goes either in another thalweg, or outside the treated<br />
zone. One can therefore count points of thalwegs’ exit by carrying out an<br />
inventory of the local extrema of the surface drained on the edge of the<br />
DTM, then go up thalwegs following this local extremum uphill (until the<br />
drained surface goes below a threshold value). Recursively, new starting<br />
points can be detected along every thalweg, looking for local extrema of<br />
the gradient of the drained surface, which correspond to the junction of<br />
an incoming drain (Rieger, 1992).<br />
Figure 3.6.6 Thalweg extracted by the highest-slope method (a) and by Jamet’s<br />
method (b) superimposed on a shaded DTM.
266 Alain Dupéret and Olivier Jamet<br />
In any case, the obtained detections can contain artefacts. The too-short<br />
drains are therefore either ignored (Rieger, 1992), or eliminated a posteriori<br />
by too-short bow suppression in the obtained network or by morphological<br />
ebarbulage.<br />
3.6.3.2 Considerations on the artefacts of the DTM<br />
The above-described methods suppose on the one hand that outside of<br />
natural basins of the landscape and the sides of the site, any point of the<br />
DTM possesses at least a neighbour of lower altitude, and on the other<br />
hand that the direction of the ground slope is defined everywhere. This<br />
last condition being equivalent to the absence of a strictly horizontal zone,<br />
one can speak of an hypothesis of strict growth of the DTM between the<br />
natural outlets of thalwegs and the summits.<br />
This hypothesis is never verified. The step of sampling data indeed limits<br />
the possibilities of representation of the present shapes on the surface, and<br />
in particular of the steep-sided valleys: so, along a thalweg crossing a valley<br />
whose width shrinks until a dimension commensurable with the sampling<br />
step, the altitude won’t be able to vary in a monotonous way and one<br />
will observe on the DTM a local minimum of altitude upstream of the<br />
steep-sided setting. In the same way, the step of quantification in altitude<br />
fixes a lower limit to the representable slopes on the DTM: any weaker<br />
slope zone will appear as a strictly horizontal zone.<br />
To these effects contingent to the discrete representation of data, some<br />
effects particular to the used source can be added: noise of acquisition on<br />
raw data generating local minima of the altitude, shortcomings of interpolation<br />
(for example, creation of horizontal zones for lack of information<br />
in the interpolation of contour lines), consideration in the representation<br />
of elements not belonging to the topographic surface (for example, vegetation,<br />
for the DTM produced by automatic image matching, that can<br />
cause an obstruction to a thalweg) . . .<br />
For grounding these artefacts, one supposes on the one hand that the<br />
ground doesn’t include a strictly horizontal zone – their process will consist<br />
then in inferring the sense of the outflow from their environment – and<br />
on the other hand that only the important basins are meaningful – this<br />
notion is specified farther.<br />
Faced with such problems, some authors have proposed solutions based<br />
on local techniques: these approaches result in giving an inertia to the<br />
outflow, which will permit it to clear zones of null slope, or even to go<br />
up the slope to come out of the artificial basins (see for example (Riazanoff<br />
et al., 1992) in the case of methods by progression). These methods don’t<br />
permit an efficient control of the restored thalwegs’ shape. We will limit<br />
the following, therefore, to methods based on a larger analysis of shapes<br />
of the relief.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Extraction of characteristic lines of the relief 267<br />
Horizontal zone process<br />
The horizontal zones can be solved in two manners: either one considers<br />
outflows as indeterminate on their extent, or one wants to define<br />
them.<br />
In the first case, the horizontal zones are treated as lakes, fed upstream<br />
by the neighbouring pixels of larger altitude (that one will call input points),<br />
and opened on the neighbouring pixels of lower altitude (that one will<br />
call output points). Practically this option means calculating the relation<br />
of outflow, no longer on the topology of the usual neighbourhood induced<br />
by the mesh of the DTM (4 or 8-connex), but on a more elaborate topology,<br />
in which the nodes may be either points (pixels) or surfaces (the horizontal<br />
zones) – while relations of the neighbourhood remain relations induced<br />
by geometry.<br />
At the end of the process of thalwegs extraction, the horizontal zones<br />
included in the tracing can be replaced by bows, whose geometry can be<br />
calculated, for example, by squeletization of these surfaces.<br />
In the second case, the choice of the outflow sense on horizontal<br />
zones is subordinated to their shape and their environment. The simplest<br />
technique consists in forcing senses of outflow from input points toward<br />
output points. For that, one can consider for example that a pixel of a<br />
flat zone is uphill of one of its neighbours if and only if its distance to<br />
the nearest exit point is greater than that of this neighbour. The distance<br />
used can be a discrete distance, calculated by dilation of the exit border,<br />
constrained by the extension of the horizontal area. This approach can be<br />
completed by a consideration of the shape of the zone: considering that<br />
the thalweg (non-observable on the DTM) must cross the centre of the<br />
zone, one can for example subordinate directions of outflow to the whole<br />
of the border of the zone, that is to say simultaneously to the distance to<br />
input points and the distance to output points as well. This last option<br />
has the advantage of also being valid if the horizontal zone does not possess<br />
an output point.<br />
Figure 3.6.7 Flow topology induced by a flat area.
268 Alain Dupéret and Olivier Jamet<br />
These types of calculation mean to accord a digital value to every<br />
pixel of the horizontal zone, which can be chosen in order to respect the<br />
required property of strict growth on the DTM. This set of values can<br />
then be assimilated to the decimal part of altitudes in calculations of<br />
slope.<br />
Process of basins<br />
One designates under the term of basin the catchment areas of the local<br />
minima of the DTM, which in most cases are correlated to artefacts. The<br />
correction of these artefacts consists in searching for one or many points<br />
of exit to the border of these catchment areas, and to force the outflow<br />
of the local minimum toward these points of exit (and therefore in the<br />
sense opposite to the slope).<br />
The position adopted by most authors consists in minimizing the incoherence<br />
of the outflow, that is to say to search for the points of exit<br />
whose height over the local minimum is the weakest possible. All basins<br />
not being necessarily artefacts, this process will generally be controlled<br />
by a threshold on this height – or denivelate, the basins whose exit are<br />
higher than the threshold being preserved (that is to say considered significant).<br />
The search for the points of exit of a basin cannot, however, be done<br />
independently of its environment: in complex situations, where numerous<br />
basins are neighbouring, the local determination of points of exit can lead<br />
to the simple fusion of neighbouring basins forming a new large-sized<br />
basin. If this process remains foreseeable while applying it recursively until<br />
disappearance of all non-meaningful basins, it can be very expensive. One<br />
has therefore rather to formulate the problem of the basin resolution like<br />
a search for a global minimum to the whole site. An economic solution –<br />
as it produces an algorithm of low complexity – consists in defining the<br />
total cost of the basin resolution as the sum of denivelates to cross in the<br />
sense inverse to the slope. Indeed, the sum of local minima altitudes being<br />
a constant of the problem, the minimization of the sum of denivelates is<br />
equivalent to the minimization of the sum of altitudes of the chosen exit<br />
points. The choice of points of exit is thus independent of the sense of the<br />
outflow, and can be operated by a simple search for the minimal weight<br />
tree in the dual graph of the basin borders, where each bow is valued by<br />
the minimal altitude of the corresponding border – that is to say the altitude<br />
of the existing lowest pass between each neighbouring basin pair.<br />
The sense of the outflow can then be calculated recursively in the obtained<br />
tree, from the known exit points (sides of the site or meaningful basins).<br />
(See Figure 3.6.8.)<br />
Once the output points are chosen, as in the case of the horizontal zones,<br />
the process of basins can give rise to a simple topological modification or<br />
to a geometric process of the DTM.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Extraction of characteristic lines of the relief 269<br />
Figure 3.6.8 Search of the basin exit: the simple choice of the lowest pass leads<br />
to the fusion of R 1 and R 2 without giving them an outlet. A global<br />
search has to be performed in the graph of neighbourhood of the<br />
basin.<br />
Among the topological processes, the two most obvious techniques<br />
consist in adding a relation of outflow between the local minimum of every<br />
basin and its pass of exit – with the drawback of producing non-connex<br />
thalwegs at the end of the process – either to reverse the sense of the<br />
outflow on pixels leading from the local minimum to the pass of exit<br />
following the steepest slope (Fairfield and Leymarie, 1991). This last technique<br />
can however be used only in the case of accumulation calculations<br />
on the outflow tree: the inversion of the sense on only one path creates<br />
indeed some cycles on the graph of outflow, and does not permit the other<br />
techniques to be used any longer.<br />
The most current geometric process consists in modifying the geometry<br />
of the DTM to construct an exit path of constant altitude, by searching for<br />
a pixel of lower altitude than the local minimum downstream of the exit<br />
pass, by levelling of the DTM on the path of steepest slope joining these<br />
two pixels (Rieger, 1992). This last solution is more adapted in the case of<br />
accumulation techniques on the entirety of the graph of outflow.<br />
Let us note finally that the order of process of the horizontal zones and<br />
the basins is not irrelevant: as the process of basins is leading to a modification<br />
of outflows on the DTM, it must evidently be applied before the<br />
process of horizontal zones.<br />
3.6.3.3 Annex results<br />
The methods presented in this section have the advantage of giving<br />
access to a description of a topography richer than the simple network of
270 Alain Dupéret and Olivier Jamet<br />
Figure 3.6.9 Watershed extracted by an outflow method: (a) DTM (shaded); (b)<br />
thalweg network; (c) watersheds (represented by various grey levels,<br />
and their boundaries in white).<br />
thalwegs. The calculation of the horizontal zones and basins, understood<br />
as artefacts of the DTM, provides an indicator of its quality. The calculation<br />
of the drained surface offers a natural hierarchy of the network<br />
of thalwegs, which can prove to be useful in numerous applications. The<br />
materialization of outflows on the whole surface permits the direct calculation<br />
of catchment areas associated to each thalweg or every portion of<br />
thalweg (for example, by simple labelling of its pixels uphill in the outflow<br />
tree). This decomposition of the surface is used extensively, in particular<br />
for the physical modelling of flood phenomena (Moore et al., 1991), but<br />
also for the cartographic generalization (Monier et al., 1996). Finally, the<br />
network of crests, subset of catchment areas borders, is also accessible by<br />
this type of technique, either that one adapts algorithms of accumulation<br />
to calculate a function of adherence to catchment areas on the same mode<br />
as the calculation of the drained surfaces, or that one uses directly calculated<br />
catchment areas to extract their borders by techniques of vectorization.<br />
3.6.4 Extraction of thalwegs from local properties<br />
This family of methods was historically the first to be used (Haralick,<br />
1983). For this algorithm and those that followed, the common idea is to<br />
search, on every pixel of the DTM, independently of results on the neighbouring<br />
pixels, to see if the current point verifies a certain relation with<br />
its environment; in which case, it receives a stamp that identifies it as a
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Extraction of characteristic lines of the relief 271<br />
point of crest or thalweg. Only some methods, considered as representative,<br />
will be presented below.<br />
3.6.4.1 By analysis of the discrete representation<br />
The surface is analysed here in a discrete way. Kim’s method (1986) is the<br />
more often evoked. The algorithm proceeds by a scan of the DTM according<br />
to X and Y in order to identify elementary geomorphologic shapes.<br />
With this method, the algorithm scans the DTM from left to right, then<br />
from top to bottom. During the sweep, characters are accorded to pixels<br />
(of the DTM) successively according to the nature of the slope that is<br />
between them. A threshold of altitude differences is applied to define the<br />
horizontal surfaces. Transitions between points are qualified according the<br />
following rules:<br />
• the passage of a point to a higher one is represented, e.g. by the character<br />
‘’;<br />
• the passage of a point to a lower one is represented by the character<br />
‘’;<br />
• the passage without change of the level of a point to a higher one is<br />
represented by the character ‘’.<br />
The interpretation of the shapes of models (images constituted by the<br />
three characters , and ) produced by two scans makes possible a<br />
first analysis, according to the scan profile, of the local morphology of the<br />
ground. The analysis relies on successions of slope characters ‘’, ‘’, ‘’,<br />
for example, if a ‘’, is followed by a ‘’, it indicates the presence of a<br />
steep-sided thalweg, while several ‘’ followed by several ‘’ followed<br />
again by one or several ‘’ could mean a large valley bottom or the flat<br />
bottom of a basin.<br />
The superposition of the two scan models, in the longitudinal and<br />
transverse sense, permits one by comparing the information, to deduce<br />
characteristics of the local morphology in the two dimensions.<br />
Thus the obtained model allows, after an ad hoc filtering, elements of<br />
the main characteristic lines network to be recovered.<br />
In certain cases, shapes are not very clean, in particular in the flat zones.<br />
In addition, the connection of the network is not always satisfactory, a<br />
certain number of segments remaining isolated.<br />
3.6.4.2 By analysis of a continuous representation<br />
The surface is assumed to be continuous, or continuous by pieces so as<br />
to provide an analytic expression of the surface. The function representative<br />
of the altitude must make it possible to give an account of the different<br />
discontinuities due to the terrain:
272 Alain Dupéret and Olivier Jamet<br />
• slope, cliffs directly bound to the altitude;<br />
• lines of slope rupture (concavity, convexity);<br />
• line of curvature rupture.<br />
On neighbourhoods of given size, the continuity of the slope can be<br />
obtained by junction of parabolic bows and that of the curvature by the<br />
utilization of cubic bow. Several methods explored this possibility; Haralick<br />
(1983) makes the analysis with the help of a model using some bi-cubic<br />
functions. The method that will be presented (Dufour, 1988) relies on the<br />
hypothesis that the ground can be modelled with the help of a Taylor<br />
series, which will be limited here to the order 2, which permits one to<br />
describe a sufficiently large number of phenomena.<br />
Z H(X,Y) H(X 0 ,Y 0 ) ax by (cx 2 2dxy ey 2 ) ,<br />
where x X X 0 and y Y Y 0 are considered small.<br />
Formulae of determination of coefficients of the polynomial depend on<br />
the chosen mesh. The square configuration will be used here and the coefficients,<br />
calculated by the least squares method are expressed in the<br />
following manner (Dupéret, 1989):<br />
H 0 5H 0<br />
9<br />
2<br />
9 (H 2 H 4 H 6 H 8) 1<br />
9 (H 1 H 3 H 5 H 7)<br />
a 1<br />
6 (H 1 H 7 H 8 H 3 H 4 H 5 )<br />
b 1<br />
6 (H 1 H 2 H 3 H 5 H 6 H 7)<br />
c 1<br />
3 (H 8 H 4) 2<br />
3 (H 2 H 6) 1<br />
3 H 1 H 3 H 5 H 7 2<br />
3 H 0<br />
d 1<br />
4 (H 1 H 5 H 3 H 7)<br />
e 1<br />
3 (H 2 H 6) 2<br />
3 (H 8 H 4) 1<br />
3 H 1 H 3 H 5 H 7 2<br />
,<br />
3 H0 where H i is defined according to Figure 3.6.10.<br />
The horizontal curvature is a very interesting variable, which uses all<br />
first and second derivatives of the altitude function. It takes very strong<br />
negative values in thalwegs, remains weak in the regular sides and becomes<br />
very strongly positive along the crests.<br />
The horizontal curvature at point (x, y) (0, 0) of contour lines is<br />
expressed according to the coefficients above:<br />
1<br />
2
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3<br />
4<br />
2<br />
0<br />
5 6 7<br />
Extraction of characteristic lines of the relief 273<br />
1<br />
8<br />
Figure 3.6.10 Numerotation of the points of the neighbourhood, of altitude<br />
H 0, H 1 ... H 8.<br />
N .<br />
The same principle is applicable to the calculus of all variable geomorphometric<br />
types expressed according to derivatives of the altitude function:<br />
slope, curvature of level lines, curvature of slope lines, Laplacian, quadratic<br />
mean curvature, total curvature, etc.<br />
2abd cb2 ea2 (a2 b2 ) 3/2<br />
?<br />
R2<br />
?<br />
R1<br />
Figure 3.6.11 Curvature radius R 1/ N in two points of the ground.
Figure 3.6.13<br />
Figure 3.6.12<br />
Representation of N<br />
(from the darkest to the<br />
lightest grey, the thalwegs,<br />
the crests, then the<br />
intermediate zones) with<br />
the superimposition of<br />
contour lines.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
3.6.4.3 Synthesis<br />
The search for points of the characteristic lines independently from each<br />
other eludes the very notion of characteristic line. If the acquisition<br />
mode and the density are appropriated, then these methods can provide<br />
a posteriori these lines. However, in the main, the extracted network is<br />
non-connex with sometimes segments of non-negligible thickness. Even if<br />
they are very robust, these algorithms very rarely restore completely<br />
the logic of a hydrographic network. The identified points often form<br />
dispersed segments, therefore non-connex, and these sets of points have a<br />
variable thickness.<br />
Some mixed methods were also set up. Thus, the algorithm of Ichoku<br />
(1993) proceeds by combination of Kim’s and Riazanoff’s algorithms and<br />
provides a hierarchized and reasonably connected network. Basins are eliminated<br />
according to the principle of the inversion of the sense of the<br />
drainage. Although the general connection of the network is superior to<br />
the two basic methods used, some parasitic networks subsist.<br />
3.6.5 Numerous applications<br />
Extraction of characteristic lines of the relief 275<br />
Examples that follow are by no means exhaustive. They are mentioned<br />
merely to show that some varied applications can be built on data produced<br />
by the above mentioned methods. Some do not require the connection of<br />
segments that compose the network. The mastery of the representation<br />
of characteristic lines network is the basis of various measures on watersheds.<br />
3.6.5.1 The hierarchization of a hydrographic network<br />
The networks of thalwegs produced by different methods can allow the<br />
hydrologists to build a hierarchical classification of segments composing<br />
the detected network. Two methods among all those possible are presented<br />
here.<br />
The first use by Riazanoff is introduced by Shreve (1967). A segment is<br />
a part of the active network of a point source to a confluence, or of a<br />
confluence to another confluence. Every segment possesses a hierarchical<br />
value that is the sum of the hierarchical values of immediately upstream<br />
segments, segments that are the most upstream (descended of a source)<br />
receiving the hierarchical value 1. (See Figure 3.6.14.)<br />
A second method, due to Horton (1945) and improved subsequently,<br />
establishes the hierarchy in the following way:<br />
• rivers not having even received an affluent are of order 1;<br />
• rivers having received at least a river of order 1, therefore 2 rivers of<br />
order 1 are of order 2;
276 Alain Dupéret and Olivier Jamet<br />
1<br />
1<br />
1<br />
3<br />
1<br />
Figure 3.6.14 Example of network in the classification of Shreve.<br />
1<br />
1<br />
1<br />
2<br />
2<br />
1<br />
Figure 3.6.15 A similar network in the classification of Horton.<br />
3<br />
2<br />
• rivers having received at least a river of order 2, therefore 2 rivers of<br />
order 2 are of order 3;<br />
• and so on.<br />
(See Figure 3.6.15.) The final order of the outlet river of a catchment area<br />
has a comparable value from a basin to the other, provided that they result<br />
from surveys of the same scale.<br />
2<br />
2<br />
1<br />
1<br />
6<br />
3<br />
1<br />
1
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Extraction of characteristic lines of the relief 277<br />
It is then possible for the user to establish relations of comparison between<br />
the order, the designation of the river and the surface of catchment areas.<br />
For example, the denomination ‘large river’ will be reserved to segments<br />
of order 7 and 8, with associated catchment areas of a surface between<br />
10,000 and 100,000 km 2 .<br />
With the help of this hierarchization of networks, the user can also use<br />
thresholds on the order of segments in order to limit some parts of the<br />
drainage network.<br />
3.6.5.2 The cartographic smoothing<br />
Despite all precautions, the filtered DTM presents softer ground shapes in<br />
comparison to what they should be. To limit the deterioration of terrain<br />
shapes, one may therefore use a cartographic smoothing process that has<br />
to determine as automatically as possible the zones of the DTM that correspond<br />
to the characteristic lines of the ground.<br />
For example, the program used in IGN-F calculates a digital model of<br />
horizontal curvature in the DTM (performing a classification in five<br />
domains) with homogeneous geomorphometric properties describing highly<br />
convex zones (crests) up to strongly concave zones (thalwegs).<br />
Parameters of adapted smoothing can then be applied in an uniform<br />
way on each of its zones to get the contour lines smoother, but preserving<br />
a good description of the terrain shapes. (See Figure 3.6.16.)<br />
Figure 3.6.16 Comparison of level curves restored by hand (in grey) and<br />
cartographically smoothed (in dark grey).
278 Alain Dupéret and Olivier Jamet<br />
3.6.5.3 Automatic cut-off of watersheds<br />
As an indispensable component to the rational management of digital data<br />
in GIS, the limits of watersheds has for a long time been digitized by hand<br />
before being inserted digitally. The first studies in automation started in<br />
1975. One of the first methods (Collins, 1975) proposes to classify all<br />
points of the DTM by increasing order of altitude. The lowest point is an<br />
outlet of the watersheds; if the second lowest point is not connex to it, it<br />
means therefore that it belongs to another basin. The provided results are<br />
good but are unstable with a real DTM where objects on the ground, the<br />
plane zones and irregularities of thalwegs’ profiles destabilize the procedure<br />
of detection.<br />
Since then numerous methods have been studied, trying to extract from<br />
the DTM the points presenting the property of adherence to a line of crest.<br />
The different strategies used to make the set of points connex and of unit<br />
thickness have most of the time stumbled on the discontinuous characters<br />
of lines obtained and the problem of local minima. The assimilation of<br />
the main and secondary crest lines in methods of non-hierarchical detection<br />
puts the problem of identification of the real lines of water sharing.<br />
The theory of the mathematical morphology constituted an efficient and<br />
original contribution to this type of survey while permitting an automatic<br />
method of catchment areas cut-off to be finalized (Soille, 1988).<br />
3.6.5.4 Manual editing of DTM<br />
The characteristic line extraction as it has just been presented is often<br />
disrupted in practice by the presence of micro-reliefs, artefacts of calculation,<br />
as small crests, passes or aberrant basins, without physical reality.<br />
The elimination of these topographic anomalies can be performed by<br />
replacing the series of elevations, in the considered thalweg, by a new set<br />
of data, obtained by a controlled diminution of the values. It is easy but<br />
poses the problem of other altitude modifications to the neighbourhood<br />
of the thalweg. The step of interactive manual correction proves to be, for<br />
the meantime, still indispensable. On its duration depends the economic<br />
opportunity to make these corrections by hand. At best, three DTM editors<br />
are accessible in some software:<br />
• A punctual editor, intended to modify only one altitude at a time (used<br />
rarely).<br />
• A linear editor working in such a way that the operator indicates interactively<br />
a 2D or 3D axis with a chosen influence distance. Inside the<br />
concerned zone, altitudes are resampled according to hypotheses made<br />
on the desired profile of ground perpendicularly to the axis indicated<br />
by the operator (profile in U, in V, linear, bulldozer . . .). This editor<br />
is useful for suppressing altimetric prints of hedges in the DTM, or<br />
when characteristic lines and/or a hydrographic network is available.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Extraction of characteristic lines of the relief 279<br />
Figure 3.6.17 In black: example of linear object that can give rise to updating of<br />
the profile of the ground in the DTM among the four possibilities<br />
and distance of interpolation (D) to the axis of the object: uniform<br />
interpolation, in U, in V or some in bulldozer. Surfacic processes<br />
can be applied to objects systematically known by their perimeter:<br />
a clear hydrographic surface (in the bottom-left corner, surrounded<br />
by a clear line) can be given a rigorously constant altitude, a forest<br />
(right, surrounded by a clear line) can receive a bias corresponding<br />
to the mean height of the trees.<br />
• A surfacic editor that allows the operator to select a process to apply<br />
to altitudes situated inside a polygon: various resamplings, force to a<br />
constant level, application of a bias, parametrable filtering of the<br />
objects over the ground . . . .<br />
The manual correction operations are shorter and more efficient if the<br />
DTM quality is good. At this step, in spite of everything, the user will not<br />
avoid some topographic anomalies, like the thalweg perched in the convex<br />
zone, the crest enclosed in a concave zone or the delta of a river in which<br />
the stream passes from convergent to divergent zones.<br />
3.6.5.5 Toward a partition of the ground in homogeneous<br />
zones?<br />
The different adopted definitions lead to a partition of the domain in<br />
catchment basins that must be completed to achieve morphologically a<br />
separation in homogeneous domains.<br />
The decomposition of the ground in homogeneous zones under shape<br />
of curvilinear triangles is foreseeable. It represents immediately a finality
280 Alain Dupéret and Olivier Jamet<br />
for the topographer and a way toward a rational generalization of the<br />
relief. The junction between two neighbouring triangles takes place along<br />
a sinuous curve corresponding to singularities of the function altitude in<br />
general. Such a model is attractive because it inserts without difficulties<br />
delicate topographic situations such as:<br />
• valleys with a low slope with a sinuous course, and<br />
• rocky-toothed cliffs,<br />
which required a decomposition of the domain into numerous facets.<br />
Every curvilinear triangle includes a parametric representation with a<br />
limited number of parameters (e.g. 30 coefficients for a parametric system<br />
of cubic type). The value taken by these functions in the border of zones<br />
makes it possible to recover the curvilinear character of separation lines.<br />
In the curvilinear domain (ABC), X, Y and H are represented by cubic<br />
functions of the barycentric coordinates of the ABC triangle. Cubic functions<br />
X(, , ), Y(, , ), H(, , ) include each ten independent<br />
parameters that can be defined by values of (X, Y, H) to the three summits,<br />
as well as in two points on every curvilinear side and one central point.<br />
and<br />
<br />
area (MBC)<br />
area (ABC)<br />
1<br />
.<br />
area (MAC)<br />
area (ABC)<br />
area (MBA)<br />
area (ABC)<br />
Separation lines are to be looked for giving priority to where there are<br />
obvious discontinuities:<br />
• of the function altitude (slope, cliff);<br />
• of the derivative of the slope (crest, steep-sided thalweg);<br />
• of the curvature (slope rupture).<br />
The convenient exploitation of such a model presents various difficulties<br />
whose major problem is the reliable automatic determination of lines of<br />
homogeneous zone separation.<br />
3.6.5.6 Complementary approach<br />
The automated extraction of the characteristic line networks is of course<br />
a privileged way of knowing the surface of the land, which comes in<br />
complement to other approaches, for example: the determination of surface<br />
envelopes (Carvalho, 1995). A surface envelope is a general, schematic and<br />
theoretical model of the relief that makes it possible to recover, in certain<br />
cases, the general lines of a situation of maturity of the relief, and to
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
analyse the links between the present and old topographies. The surface<br />
envelope relies on the high points of the relief of a region, so as to eliminate<br />
irregularities of the topographic surface coming from the linear<br />
erosion of rivers. Indicators of surface thickness either eroded or erodable<br />
can be deduced from differences between DTM and the surface envelope<br />
to provide some useful elements to the evaluation of erosion balances.<br />
The automatic detection of summit points is considered here therefore<br />
like a tool serving the development of a numeric representation of surface<br />
envelopes while giving points of passage. It is followed by a phase of<br />
construction of the surface. The practice shows that the automation<br />
of this determination is justified when the DTM mesh is adapted to the<br />
land. The simultaneous knowledge of networks of thalwegs and surface<br />
envelopes allows the imperfections of the thalwegs network in relation<br />
to the summit surface envelopes to be determined (direction of the thalwegs<br />
network not compliant with the lines of steepest slope of the surface<br />
envelopes).<br />
References<br />
Extraction of characteristic lines of the relief 281<br />
Carvalho J. (1995) Extraction automatique d’informations géomorphométriques<br />
(réseaux et surfaces enveloppes) à partir de modèles numériques de terrain. Ph.D.<br />
Thesis, Université de Paris 7.<br />
Chorowicz J., Parrot J.F., Taud H., Hakdaoui M., Rudant J.P., Rouis T. (1995)<br />
Automated pattern-recognition of geomorphic features from DEMs and satellites<br />
images Z, Geomorph. N.F. (Berlin), Suppl. Bd101, pp. 69–84.<br />
Collins S. (1975) Terrain parameters directly from a digital terrain model, Canadian<br />
Surveyor, vol. 29, no. 5, pp. 507–518.<br />
Depraetère C., Moniod F. (1991) Contribution des modèles numériques de terrain<br />
à la simulation des écoulements dans un réseau hydrographique, Hydrol.<br />
Continent. vol. 6, no. 1, pp. 29–53.<br />
Dufour H.M. (1977) Représentation d’une fonction par une somme de fonctions<br />
translatées, Bulletin d’information de l’IGN, no. 33, pp. 10–36.<br />
Dufour H.M. (1983) Eléments remarquables du relief – définitions numériques utilisables,<br />
Bulletin du comité français de cartographie, no. 95, pp. 57–90.<br />
Dufour H.M. (1988) Quelques idées générales concernant l’établissement et<br />
l’amélioration des Modèles Numériques de Terrain, Bulletin d’information de<br />
l’IGN, no. 58, pp. 3–18.<br />
Dupéret A. (1989) Contribution des MNT à la géomorphométrie, Rapport de<br />
stage, DEA SIG, IGN – IMAGEO.<br />
Fairfield J., Leymarie P. (1991) Drainage network from grid digital elevation models,<br />
Water Resources Research, vol. 27, no. 5, May, pp. 709–717.<br />
Freeman T.G. (1991) Calculating catchment area with divergent flow based on<br />
regular grid, Computer and Geosciences, vol. 17, no. 3, pp. 413–422.<br />
Haralick R.M. (1983) Ridges and valleys on digital images, Computer Vision,<br />
Graphics and Image Processing 22, pp. 28–38.<br />
Horton R.F. (1945) Erosional development of streams and their drainage basins:<br />
hydrophysical approach to quantitative morphology, Geological Society of<br />
America Bulletin, vol. 56, pp. 275–370.
282 Alain Dupéret and Olivier Jamet<br />
Ichoku C. (1993) Méthodes automatiques pour l’analyse et la reconnaissance des systèmes<br />
d’écoulement en surface et dans le sous sol. Thesis, Université de Paris 6.<br />
Jamet O. (1994) Extraction du réseau de thalwegs sur un MNT, Bulletin d’information<br />
de l’IGN, no. 64, pp. 11–18.<br />
Kim Y.J. (1986) Reconnaissance de formes géomorphologiques et géologiques à<br />
partir de modèles numériques de terrain pour l’exploitation de données stéréoscopiques<br />
Spot. Ph.D. Thesis, Université de Paris 6.<br />
Le Roux D. (1992) Contrôle de la cohérence d’un réseau hydrographique avec un<br />
modèle numérique de terrain, Rapport de stage, Laboratoire COGIT, IGN, Saint-<br />
Mandé, France.<br />
Le Roux D. (1993) Modélisation des écoulements sur un modèle numérique de<br />
terrain – Applications aux crues et inondations du 22/09/89 à Vaison la Romaine,<br />
Engineer End of Study Memoir, ESGT, Le Mans, France.<br />
Monier P., Beauvillain E., Jamet O. (1996) Extraction d’éléments caractéristiques<br />
pour une généralisation automatique du relief, Revue Internationale de Géomatique,<br />
vol. 6, no. 2–3, pp. 191–201.<br />
Moore I.D., Grayson R.B., Ladson A.R. (1991) <strong>Digital</strong> terrain modelling: a review<br />
of hydrological, geomorphological and biological applications, Hydrological<br />
processes, vol. 5, pp. 3–30.<br />
Riazanoff S. (1989) Extraction et analyse automatiques de réseaux à partir de<br />
modèles numériques de terrain. Contribution à l’analyse d’images de télédétection.<br />
Ph.D. Thesis, Université de Paris 7.<br />
Riazanoff S., Julien P., Cervelle B., Chorowicz J. (1992) Extraction et analyse<br />
automatiques d’un réseau hiérarchisé de thalwegs. Application à un modèle<br />
numérique de terrain dérivé d’un couple stéréoscopique SPOT, International<br />
Journal of Remote Sensing, vol. 13, pp. 367–364.<br />
Rieger W. (1992) Automated river line and catchment area extraction from DEM<br />
data, ISPRS Congress Commission IV, Washington DC, August, pp. 642–648.<br />
Shreve R.L. (1967) Infinite topologically random channel networks, Journal of<br />
geology (Chicago), vol. 75, pp. 178–186.<br />
Soille P. (1988) Modèles numériques de terrain et morphologie mathématique:<br />
délimitation automatique de bassins versants, Engineer End of Study Memoir in<br />
agronomy, orientation ‘génie rural’, Université catholique de Louvain la Neuve,<br />
Belgium.<br />
3.7 FROM THE AERIAL IMAGE TO ORTHOPHOTOGRAPHY:<br />
DIFFERENT LEVELS OF RECTIFICATION<br />
Michel Kasser, Laurent Polidori<br />
3.7.1 Introduction<br />
The very principle of an image acquisition, which means a conical perspective<br />
for specialists, implies that a photographic image is not generally<br />
superimposable to a map. It is partly due to the existing reliefs in the<br />
observed object, and to the fact that even for an object that would be<br />
rigorously plane, the optical axis has no reason to be precisely perpendicular<br />
to this plane. For a very long time, techniques producing the
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
controlled distortions of photographs have been developed in order to give<br />
these images a comparable metrics to that of a map. These techniques can<br />
be conceived at very elaborate levels, according to the requested quality.<br />
Rigorously speaking, only when the correction is perfect is the product<br />
obtained called orthophotography, but there are many intermediate solutions,<br />
that one uses to get rectified images. Here we aim to clarify the<br />
different levels possible, which have some very different costs, and corresponding<br />
large differences of precision.<br />
3.7.2 What is an orthophotography?<br />
From the aerial image to orthophotography 283<br />
The orthophotography is a picture (generally aerial) that has been geometrically<br />
rectified to make it superimposable in any place on a map, possibly<br />
with enriched graphic additions. These additions can originate either from<br />
external information (administrative limits, toponyms, etc.), or from the<br />
interpretation of the image itself (in order to ease its use: drawing of roads,<br />
of building contours, etc.). If one compares it with classic cartography,<br />
the orthophotography differs by the absence (that can be total) of most<br />
phases of interpretation and drawing. As such tasks always require an<br />
important human added value, the possibility of suppressing them permits<br />
therefore an almost total automation of the process: it is through the<br />
considerable reduction of costs and delays obtained that digital orthophotography<br />
became a product more and more current, capable of completely<br />
replacing the traditional cartography.<br />
The process of manufacture necessarily implies knowing two data sets:<br />
parameters of image acquisition (orientation of the image, spatial position<br />
of the camera) and the relief (described generally by a DTM, digital terrain<br />
model). It is therefore clear that the precision of the final product depends<br />
directly on the quality of these two data sets.<br />
Let us note here that one knows how to get a DTM currently by automatic<br />
image matching of images. It is more or less today the only operation that<br />
may be completely automated in digital photogrammetry, as the other tools<br />
that tomorrow may be accessible (automatic drawing of roads, extraction of<br />
buildings, etc.) currently require the supervision of a human operator.<br />
The user will sometimes not need a very advanced cartographic quality,<br />
only a fairly constant mean scale, for example within 10 per cent, which<br />
is not compatible with the raw photo but may be performed without<br />
rigorous photogrammetric process. For example, commercial software for<br />
image manipulation (advertisement, editing, retouching of amateur photographs,<br />
etc.) may be used the proper way: one will search for some<br />
identifiable points of known coordinates, and one will distort the image<br />
in order to oblige it to respect their positions. Certainly it is not at all an<br />
orthophotography, but it may provide some facilities and will often be less<br />
expensive. But it is necessary to avoid using the same denomination: one<br />
will speak, for example, of rectified images. The term ‘orthorectified’ image
284 Michel Kasser and Laurent Polidori<br />
or orthophotography would be an abuse here. We will use the generic<br />
term ‘rectified’ image for any image that underwent a process that distorts<br />
it geometrically to bring it closer to an orthophotography. This rectification<br />
can therefore be performed at various levels: so the orthophotography<br />
represents the highest possible level concerning rectification.<br />
Otherwise, most users need numerous images to cover a given area, which<br />
requires a ‘mosaiquage’, meaning that it is necessary to process the links<br />
between images so that one will no longer notice them, in terms of geometry<br />
(breakage in linear elements) and of radiometry as well (differences<br />
of grey level, or of hue, for one given object on two nearby images). For<br />
the geometric aspects, the previous considerations will help to get an idea<br />
of the nature of the problems susceptible of being met. On the other hand,<br />
for radiometric aspects, the physical phenomena originating them are<br />
numerous: chemical process of images, difference of quantity of light<br />
distributed by the same object in different directions (bidirectional reflectance),<br />
images possibly acquired at different dates, or merely at different<br />
instants (and therefore lighting), etc.<br />
Let us add again a few elements of terminology used in the spatial images<br />
process:<br />
• One can say that a image is ‘georeferenced’ if a geometric model is<br />
defined, permitting one to calculate the ground coordinates of every<br />
point of the image. For example, a grid of distortions will be provided<br />
with the image, this remaining raw otherwise.<br />
• An image is said to be ‘geocoded’ if it has been already more or less<br />
rectified (with, in consequence, a better scale constancy).<br />
3.7.3 The market for orthophotography today<br />
One will have understood that all these photographic products have in<br />
common to leave the work of interpretation to the final user. Experience<br />
shows besides that, if most users of geographical information have in<br />
general sometimes important difficulties in reading and exploiting a map,<br />
on the other hand practically everybody will be able, even without any<br />
technical culture, to understand and to interpret the full content of a photographic<br />
document. Multiple cadastral investigation examples in rural Africa<br />
can be found in support of this observation.<br />
This ease of use makes it in fact an excellent support of communication,<br />
or even of negotiation, in particular for decision making related to regional<br />
development (individual citizens, elected people, administrations, etc.).<br />
Otherwise, as mentioned previously, the development of rectified images<br />
is capable of a very high level of automation, which has considerably<br />
decreased the production costs for some years. One even foresees shortly an<br />
entirely automatic production possibility, with very common computers of<br />
PC-type. One will thus have access to a product hardly more expensive than
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
the image of origin (this is more or less true as far as the DTM is available),<br />
and that will be able to be practically orthorectified without additional cost.<br />
Besides, it could become of little use to produce badly rectified images.<br />
Let us note again that the recent evolutions of the PC permit comfortable<br />
manipulations of digital images, even the very voluminous and numerous<br />
ones, so that the orthophotography often became the basic layer of urban<br />
GIS, the only one that may be updated regularly.<br />
There are therefore many reasons to anticipate a major development of<br />
the market of the digital orthophotography, and it will not necessarily be<br />
the same for imperfectly rectified images.<br />
Besides, several countries have undertaken or even finished and maintain,<br />
a national coverage in digital orthophotographies.<br />
3.7.4 The different levels of rectification<br />
From the aerial image to orthophotography 285<br />
Given that heterogeneous metric quality products discussed previously<br />
coexist, it appears desirable to clarify concepts of the used terminology.<br />
There are only five applicable quality criteria on the subject, if we look<br />
at the customer’s needs:<br />
1 Geodetic criterion: precision of the correspondence between coordinates<br />
of objects in the official national reference frame and in the<br />
rectified image (if coordinates are displayed in the document): so, if<br />
the system of coordinates is not sufficiently known (for lack of links<br />
to the local geodesy for example), the coordinates that the user can<br />
extract from the image will be in error by a constant value, which can<br />
be very significant. This type of error, for certain applications, has no<br />
importance, but will be bothersome in others (e.g. for somebody<br />
performing further work with GPS).<br />
2 Scale quality criterion: stability of the image scale, or more precisely<br />
evaluation of the scale gap between the image and the corresponding<br />
theoretical map (whose scale is not precisely constant according to the<br />
system of projection used). One can speak of a scale stability within<br />
10 per cent (coarse process), and up to 0.01 per cent (the most precise<br />
processes). This criterion is certainly the most important of all. Let’s<br />
not forget here that the precision of the DTM (a concept otherwise<br />
difficult to define exactly) appears directly in this criterion: if the DTM<br />
is false, the rectification will necessarily be the same!<br />
3 Object scale criterion: in relation with the criterion (5) below, does<br />
the criterion (2) apply only to the ground, or also to objects above<br />
the ground? This point is currently important because the DTM only<br />
concerns the ground, and removes, as much as possible, these objects<br />
(buildings, trees, etc., see §3.3).<br />
4 Radiometric process criterion: concerning the visual quality of the rectified<br />
image, we must note that all resampling will imply a certain
286 Michel Kasser and Laurent Polidori<br />
deterioration of the contour cleanness. The visual quality of the image<br />
assembly includes several more or less appreciable aspects: specular<br />
reflections on water surfaces (whose suppression sometimes implies<br />
heavy manual work), balancing of hues on links (that may require a<br />
human intervention for a good result), etc.<br />
5 Building process criterion: in the case of the urban zones, the presence<br />
of large buildings creates specific difficulties often implying a<br />
manual intervention (digital model of the relief is incomplete because<br />
of zones that are not seen in stereo, hidden zones where images’ sources<br />
do not give any information, etc.). In this criterion, two important<br />
technical aspects are the value of the focal length of the camera used<br />
(the more important it is, the less the image acquired is usable to get<br />
the DTM, so that it will be necessary with the use of a long focal<br />
distance to have the DTM already done elsewhere, but the more the<br />
problems of building facades will be reduced), and the overlap between<br />
the exploited images (if the overlap is very significant, it allows one<br />
to use only the central zones of images, where there are little or even<br />
no problems with buildings).<br />
For these last two aspects, the operator’s interventions can be more important.<br />
The final quality is thus quite difficult to specify, and it implies some<br />
very variable and sometimes major costs.<br />
Finally one will be able to refer also to process specifications in use in<br />
the spatial imagery. Levels of SPOT images processes are typical examples<br />
of what is currently proposed:<br />
• Level 1: raw image.<br />
• Level 2: rectified image with the exclusive use of image acquisition<br />
parameters, but without information on the relief (product being sufficient<br />
in zones of plane for some applications). Such a product will be<br />
accessible when one obtains the digital images with their orientation<br />
and localization parameters.<br />
• Level 3: image rectified by using image acquisition parameters and a<br />
DTM. The result will be therefore an orthophotography.<br />
But currently, no terminology has been dedicated by the use to distinguish<br />
the different qualities of process. It would be most helpful if the rectified<br />
products suppliers systematically provide evaluations on the five criteria<br />
above. For the development of this sector of activities, it is indeed not<br />
desirable that the present confusion persists. For example criteria (2) and<br />
(3), although quite essential, are almost never specified today.<br />
3.7.5 Technical factors seen by the operator<br />
Let us now make a synthesis of elements of the technical process concerning<br />
the operator in the final quality of a rectified image:
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
• The DTM can be obtained directly from the exploited stereoscopic<br />
cover and the corresponding stereopreparation. Criteria: precision of<br />
the stereopreparation, quality of the altimetric restitution. If the DTM<br />
is provided by the customer, or obtained in available external databases,<br />
one will observe a separation of responsibilities between the<br />
operator and the supplier of the DTM, which can lead to a complete<br />
removal of responsibility from the operator! Criterion: precision of<br />
altitudes of the DTM.<br />
• Parameters of image acquisition are, in all present processes, unavoidable<br />
by-products of the photogrammetric part of the process<br />
(aerotriangulation). Criterion: are the parameters of image acquisition<br />
used?<br />
• The knowledge of the reference frame is not a simple problem. It even<br />
occurs in certain regions of the world that one doesn’t have access to<br />
it at all. Without reaching this point, there is again there a sharing of<br />
responsibilities between the operator and the supplier of reference<br />
points, capable of removing the responsibility from the operator.<br />
Otherwise, the altimetry being generally referenced on the geoid and<br />
planimetric coordinates being purely geometric (referenced on an ellipsoid),<br />
a large confusion often reigns in altimetric reference systems<br />
used (national levelling network, GPS measures, etc.). Besides, traditional<br />
national planimetric networks have, at the time of a generalized<br />
use of the GPS, important errors often hardly known: one cannot reference<br />
an image rectified in such a system without adjusting the error<br />
model locally, and it implies ground measurements. Criterion: knowledge<br />
of the planimetric and altimetric reference frames.<br />
• The human intervention is more critical to reach the desired quality<br />
(in particular visual): study of reference points, geometric continuity<br />
across the links, radiometric continuity, process of buildings, etc.<br />
Criteria: qualifications, practice and care of the operator.<br />
It is finally necessary to put users on guard against the fact that more and<br />
more commercial software proposes the functionality of geometric distortion<br />
of an image that can look like a rectification of the image. It is due<br />
to the explosion of the market of the digital photograph (warping,<br />
morphing, etc.). As we saw previously, these processes that are completely<br />
unaware of the geometric origin of the treated problem can only provide<br />
some mediocre results. At a time where a rigorous orthorectification is<br />
becoming almost entirely automatic and therefore financially painless, we<br />
can only encourage the users to take care with such products.<br />
3.7.6 Conclusion<br />
From the aerial image to orthophotography 287<br />
The customer sometimes accepts too easily the available product limits for<br />
lack of a sufficient culture in geographical information. A clarification of
288 Michel Kasser and Laurent Polidori<br />
the nomenclature of products and the underlying concepts is absolutely<br />
necessary if one wants to enjoy the expansion of this market, which must<br />
improve on healthier bases. Otherwise a certain ‘democratization’ of<br />
orthorectified images appears indisputable today. If beneficiaries of service<br />
in this matter don’t come from the geomatics technical environment, and<br />
don’t have a sufficient culture concerning geographical information (in<br />
particular concerning photogrammetry), it may be anticipated that the<br />
available products will for a long time present a quality level much lower<br />
than what is possible, this despite the perfectly good faith of the operator,<br />
and for the same price as a rigorous process . . .<br />
3.8 PRODUCTION OF DIGITAL ORTHOPHOTOGRAPHIES<br />
Didier Boldo<br />
3.8.1 Introduction<br />
In theory, a line of orthoimages production can be analysed in very simple<br />
modules. It uses in input the images, the geometry of the image acquisition<br />
and a digital surface model (DSM). From these data one may build<br />
the orthoimages covering the whole zone. One calculates a mosaic, does<br />
a radiometric balancing and then archives the results. (See Figure 3.8.1.)<br />
Images can originate from numerous sources. Generally, one uses very<br />
precisely silver halide scanned images (pixel of 21 to 28 microns). Their<br />
main inconvenience is their lack of radiometric consistency that creates<br />
problems at the time of the orthoimages mosaics calculation. These images<br />
generally have some very important sizes (typically of the order of 300<br />
Mo for colour images), which requires computer power for their manipulation.<br />
If these data are used, the line of production must foresee some<br />
professional scanners, as well as the companion equipment.<br />
But new types of data become available: images originating from digital<br />
cameras, from linear CCD aerial sensors and high-resolution satellite<br />
images. These data have the advantage of an excellent internal radiometric<br />
Images<br />
Geometry of image<br />
acquisition<br />
<strong>Digital</strong> surface<br />
model<br />
Orthoimages<br />
computation<br />
Mosaic<br />
computation<br />
Figure 3.8.1 Production of digital orthophotographies.<br />
Radiometric<br />
balancing<br />
Archive
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
consistency, which greatly simplifies the processes of mosaic calculation.<br />
However, these processes can represent an important part of the calculation<br />
of the orthoimage, notably in terms of operator time.<br />
These images possess a geometry, produced through aero- or spatiotriangulation.<br />
These calculations allow one to determine the direction of<br />
every pixel of each image. These calculations can be integrated in the line<br />
productions. This situation has the advantage of allowing one to use algorithms<br />
in order to automate part of the work, e.g. the detection of link<br />
points (see §2.4).<br />
The last requested data is the digital surface model. In the case of classic<br />
orthoimages, it is a digital terrain model that is generally available on the<br />
shelf as outside data in most western countries. But if one wishes to make<br />
a true orthoimage, it is necessary to use a more precise model. This model<br />
can come from an aerial laser scanner, from automatic correlation, or from<br />
classic photogrammetric restitution. If these models are precise and corrected<br />
enough, they can even permit one to straighten the images of the<br />
buildings.<br />
In fact, the necessary data to manufacture orthoimages can be integrated<br />
in assembly lines of production. So the aspects of calculation of the image<br />
geometry (see §1.1), of calculation of the DSM (see §3.2), and problems<br />
of colour consistency (see §1.3) are integrated generally in the production<br />
chain.<br />
Some automatic tools for the calculation of join-up lines and for balancing<br />
radiometry are now available. They permit important gains of operator<br />
time. Balancing the radiometry of images is a complex problem, difficult<br />
to model, in particular for scanned images (mostly due to the imperfections<br />
of the photographic acquisition, see §1.8). Problems of colour are due to<br />
movements of the sun and variations of the angle optical axis–sun. It generates<br />
problems of hot spot and of specular reflections. One zone has different<br />
aspects therefore according to the chosen image.<br />
3.8.2 Line of join up<br />
Production of digital orthophotographies 289<br />
The automatic calculation of join-up lines on images of good radiometric<br />
quality is a relatively simple problem. Let’s place us in the zone of the<br />
two orthoimages superposition. In this zone, every ground point possesses<br />
two representations: one in each orthoimage. The line of join-up is merely<br />
a path joining two opposite sides of this zone. One wants this line to be<br />
as invisible as possible. For that, one is going to assign a cost value to<br />
every point of the recovery zone. This value represents the ‘visibility’ of a<br />
join-up line if it passes by this point. A measure can be the difference<br />
between the grey level of pixels. Another measure can be the presence of<br />
a contrast between the two images. The line of join-up will then be the<br />
minimum cost path. (See Figure 3.8.2.)
290 Didier Boldo<br />
Figure 3.8.2 Automatic join-up example between four digital images.<br />
3.8.3 Radiometric balancing<br />
As is clearly visible on Figure 3.8.3 (colour section), radiometric differences<br />
between images, especially scanned images, can be very important.<br />
They may have many causes, such as the chemical aspects of the development,<br />
the adjustment of the scanner and the movements of the sun.<br />
Misleading differences between images can be very bothersome. Besides,<br />
these differences can put in failure the algorithm of calculation of join-up<br />
line presented above. For scanned images, the problem of modelling the<br />
difference is very complex, or even impossible. The used methods are therefore<br />
generally empirical. They use parameters that must be adjusted by<br />
hand and require a certain experience.<br />
One of the essential problems is the volume of data. Thus, the regular<br />
cover (every 5 years) of French territory performed by IGN-F by a colour<br />
orthophotography of ground pixel 50 cm represents 1,200 Go of data. It<br />
is necessary to bring in at least as much data for images, then for overlap<br />
zones as well as auxiliary data. It represents 3 Tera-Bytes (4,800 CD-<br />
ROM) of data that must transit production lines and be archived every<br />
year. Material and software solutions exist, but are relatively heavy to<br />
operate. Obviously this part of the problem will progressively disappear<br />
with the evolution of informatics.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 3.8.3 Scanned images before balancing.<br />
Figure 3.8.4 Scanned images after balancing.<br />
3.8.4 Conclusion<br />
Production of digital orthophotographies 291<br />
Having examined the bases of the orthophotography process, we will see<br />
in the next section (§3.9) how to perform the practical production itself.<br />
An important point must nevertheless be raised now: as mentioned in<br />
§3.7 the production of orthophotographies is quite different if we are in<br />
urban or in rural zones. In rural areas, the main quality aspect will come<br />
from the correct mosaicking (particularly the problem of ‘hot spot’),<br />
nobody will notice the way the trees or the fences are represented. In urban<br />
areas, the representation of the buildings induces another set of difficulties
292 Didier Boldo<br />
already mentioned (hidden zones, very dark shadows, etc.). Thus the software<br />
used for such zones are mostly different. This must be taken into<br />
account when one thinks of a production line.<br />
3.9 PROBLEMS RELATING TO ORTHOPHOTOGRAPHY<br />
PRODUCTION<br />
Didier Moisset<br />
The orthophotography is an artificial digital picture. Completely manufactured<br />
by the computer, it looks automatically in photographs for radiometry<br />
information and borrows from the map, of automatic way, its<br />
carroyage, its bootjacks and sometimes some vector complements (contour<br />
lines, roads, site names, etc.).<br />
It is neither a photo nor a map and the customer wants it to be both<br />
as beautiful as the photo and as precise as the map. This ambiguousness<br />
can cause some problems for the producer of orthophotos. One used to<br />
say that the data size was a major problem. The advent of the fast networks<br />
and very high capacity drives allied to the increasing strength of machines<br />
makes this affirmation less and less true. Since the production is fully<br />
computerized, it is legitimate to want it completely automatic and this is<br />
the case most of the time when the image acquisition is good, the landscape<br />
simple, and the control data of good quality. One will notice that<br />
these conditions are sometimes present (particularly on the data sets that<br />
are used for software demonstrations!), but only ‘sometimes’.<br />
In a situation of mass production, conditions are sometimes very different<br />
and problems occur quickly when, in spite of the power of our machines,<br />
data don’t permit, by their nature, the association of the aesthetic quality<br />
Figure 3.9.1 Example from a real production (digital aerial camera).
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
of the photo and the precision of the map. Of course, it is when he counts<br />
in its significant cost that the volume of data manipulated becomes appreciable<br />
for the producer.<br />
The homogeneity of the radiometry<br />
Problems of orthophotography production 293<br />
Let’s imagine an automatic production chain that would only put in<br />
evidence at the time of the mosaicking that the nature of the digital pictures<br />
originating from the scanner do not allow a correct radiometric equating<br />
to be performed. (See Figure 3.9.2, colour section.)<br />
Figure 3.9.2 Example from a real production problem: the radiometric balancing<br />
on digitized pictures.
294 Didier Moisset<br />
This example from a real production is a caricature that shows us the<br />
importance of an efficient procedure of validation of scanning operations<br />
for the production chain. The replacement of a film during the aerial<br />
mission, the modification of chemical film process or a problem in the calibration<br />
of the scanner are many factors that can occur and generate locally<br />
for the mosaic bothersome shortcomings as those noted here.<br />
These problems don’t exist as soon as one uses digital cameras, but let<br />
us keep in mind that when a mission spreads in time, the modification of<br />
the landscape can become appreciable, or even intolerable for the mosaic.<br />
The interpretation of the banking of the raised structures<br />
By its very nature, the orthophoto even when it is of high quality is able<br />
to remind us that it is not a photo and sometimes the choice in the position<br />
of the join-up line has a fundamental importance for the aesthetics<br />
of the mosaic (see Figures 3.9.3 (a) and (b)).<br />
The producer must prevent this kind of fantasy and must use possibilities<br />
of modification of the automatic calculation of the join-up line by<br />
imposing points of passage if its tool possesses this functionality. If it is<br />
not the case, he should modify the join-up line by hand making it follow<br />
the characteristic lines of the landscape.<br />
This example is also caricatural, but one can imagine without difficulty<br />
the problem for the producer of the orthophoto who must define and value<br />
the level of sharpness that he needs to grant for the verification and the<br />
possible modification of join-up lines. It is therefore a costly interactive<br />
phase that is going to condition the quality of the final result.<br />
Stretch of pixels<br />
Another difficulty that one frequently meets on landscapes to strong reliefs:<br />
when the optical ray is tangent to the DTM, a displacement on the DTM<br />
doesn’t generate a displacement on the image. The image will provide the<br />
same radiometry therefore for a succession of pixels of the orthophoto.<br />
One says that the pixel is stretched. (See Figures 3.9.4 and 3.9.5, colour<br />
section.)<br />
In this case, which is not rare, the choice of the image that will provide<br />
the radiometry (therefore the position of the join-up line) is determinant<br />
for the quality of the resulting mosaic.<br />
Difficulties bound to the type of landscape<br />
Some landscapes create some difficult problems to solve. The following<br />
are two examples of landscapes whose aspect changes significantly according<br />
to the point of view from which one looks at it.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figures 3.9.3 (a) and (b) Example from a real production (area of Dijon).<br />
A correct line of join-up is not just a question of radiometry,<br />
especially in urban zones.<br />
(a)<br />
(b)
296 Didier Moisset<br />
Figure 3.9.4 Examples from the Ariège, France: problems posed by cliffs.<br />
Figure 3.9.5 Generation of stretched pixels.<br />
Sometimes, the join-up line is difficult to conceal whereas there is no<br />
alternative. It is the case in landscapes of strong reliefs where the aspect<br />
of vegetation is going to depend greatly on the angle from which one looks<br />
at it (see Figure 3.9.6, colour section).<br />
It is also the case with specular reflections well known from picture<br />
processors (Figure 3.9.7, colour section).<br />
DTM
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Figure 3.9.6 Example of radiometric difficulties (Ariège, France).<br />
Figure 3.9.7 Example of typical problems on water surfaces.
298 Didier Moisset<br />
Figure 3.9.8 Two examples from the Isère department, France.<br />
Difficulties bound to the geometry<br />
On the examples shown in Figure 3.9.8 (colour section), the radiometric<br />
homogenization (that is acceptable) is going to overlook for a non-aware<br />
eye a serious defect of geometry. It is clear in this case that the DTM and<br />
the result of the aerotriangulation that have both served to elaborate the<br />
orthophotography are not in agreement.<br />
When this observation is performed only via the examination of the<br />
mosaic (which is the final product) one easily understands the consequences<br />
for the producer who must analyse the whole chain starting with the<br />
control points of the aerotriangulation, as well as the nature of difficulties<br />
that is going to meet the person who must value the amount of work<br />
to perform at this step.<br />
But when all goes well<br />
The examples shown in Figure 3.9.9 (colour section) allow one to conclude<br />
on a positive note. When geometry and radiometry are in harmony, and<br />
in particular when the radiometry is very precise (the case of the digital<br />
cameras acquisitions), it is necessary to make the join-up line visible in<br />
the mosaic to remind one that the orthophotography is not originating<br />
from only one unique photo.<br />
By these examples, one can understand the importance that the producer<br />
of orthophotographies will pay to all operations of validation of the<br />
different steps of the process.<br />
As it uses a set of high technologies ranging from photogrammetry to<br />
the development of digital models of land and digital picture processes,<br />
the rigorous realization of an orthophotography at a time is not a simple<br />
thing, even if it presents the strong interest of being automated.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Problems of orthophotography production 299<br />
Figure 3.9.9 Two examples, using the digital aerial camera of IGN-F. Due to the<br />
linearity of the CCD, there are only very minor radiometric<br />
differences across the join-up line, which may be chosen without<br />
difficulty.
4 Metrologic applications of<br />
digital photogrammetry<br />
INTRODUCTION<br />
At the end of this book, we have a short presentation of some specific<br />
applications of digital photogrammetry, generally at closer ranges and<br />
without aerial acquisition of the images. Two main domains are presented<br />
here, the architectural applications (§4.1) and the metrologic applications<br />
(§4.2). These two domains have experienced a considerable extension in<br />
recent years, mainly because of the use of digital photogrammetry and the<br />
availability of low-cost digital photogrammetric workstations (DPW).<br />
4.1 ARCHI<strong>TEC</strong>TURAL PHOTOGRAMMETRY<br />
Pierre Grussenmeyer, Klaus Hanke, André Streilein<br />
4.1.1 Introduction<br />
Compared with aerial photogrammetry, close-range photogrammetry and<br />
particularly architectural photogrammetry isn’t limited to vertical photographs<br />
with special cameras. The methodology of terrestrial photogrammetry<br />
has changed significantly and various photographic acquisitions are<br />
widely in use.<br />
New technologies and techniques for data acquisition (CCD cameras,<br />
Photo-CD, photoscanners), data processing (computer vision), structuring<br />
and representation (CAD, simulation, animation, visualization) and archiving,<br />
retrieval and analysis (spatial information systems) are leading to novel<br />
systems, processing methods and results.<br />
The purpose of this chapter is to introduce, as part of the International<br />
Committee for Architectural <strong>Photogrammetry</strong> (CIPA), the profound changes<br />
currently stated.<br />
The improvement of methods for surveying historical monuments and<br />
sites is an important contribution to the recording and perceptual monitoring<br />
of cultural heritage, to the preservation and restoration of any
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
valuable architectural or other cultural monument, object or site, as a<br />
support to architectural, archaeological and other art-historical research.<br />
4.1.2 Strategies for image processing<br />
Architectural photogrammetry 301<br />
Close-range photogrammetry is a technique for obtaining geometric information,<br />
e.g. position, size and shape of any object, that was imaged on<br />
photos previously.<br />
To achieve a restitution of a 3D point you need the intersection between<br />
at least two rays (from photo to object point) in space or between one<br />
ray and the surface that includes this point. If more than two rays are<br />
available (the objects shows on three or more photos) a bundle solution<br />
is possible including all available measurements (on photos or even others)<br />
at the same time.<br />
These cases lead to different approaches for the photogrammetric restitution<br />
of an object.<br />
4.1.2.1 Single images<br />
A very common problem is that we know the shape and attitude of an<br />
object’s surface in space (digital surface model) but we are interested in<br />
the details on this surface (patterns, texture, additional points, etc.). In<br />
this case a single image restitution can be appropriate.<br />
With known camera parameters and exterior orientation<br />
In this case the interior orientation of the camera and the camera’s position<br />
and orientation are needed. So the points can be calculated by intersection<br />
of rays from camera to surface with the surface known for its shape and<br />
attitude.<br />
Figure 4.1.1<br />
Camera position
302 Pierre Grussenmeyer et al.<br />
Interior orientation does not mean only the calibrated focal length and<br />
the position of the principal point but also the coefficients of a polynomial<br />
to describe lens distortion (if the photo does not originate from a<br />
metric camera).<br />
If the camera position and orientation is unknown at least three control<br />
points on the object (points with known coordinates) are necessary to<br />
compute the exterior orientation (spatial resection of camera position).<br />
Without knowledge of camera parameters<br />
This is a very frequent problem in architectural photogrammetry. The<br />
shape of the surface is restricted to planes only and a minimum number<br />
of four control points in two dimensions have to be available. The relation<br />
of the object plane to the image plane is described by the projective<br />
equation of two planes:<br />
X a 1 x a 2 y a 3<br />
c 1x c 2y 1 ,<br />
Y b 1x b 2y b 3<br />
c 1x c 2y 1 ,<br />
where X and Y are the coordinates on the object’s plane, x and y the<br />
measured coordinates on the image and a i , b i , c i the eight parameters describing<br />
this projective relation.<br />
The measurement of a minimum of four control points in the single<br />
photo leads to the evaluation of these eight unknowns (a 1, a 2, a 3, . . . , c 2).<br />
As a result, the 2D coordinates of arbitrary points on this surface can<br />
be calculated using those equations. This is also true for digital images of<br />
facades. <strong>Digital</strong> image processing techniques can apply these equations for<br />
Figure 4.1.2
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 303<br />
every single pixel and thus produce an orthographic view of the object’s<br />
plane, a so-called orthophoto or orthoimage (see §3.7). (See Figure 4.1.2.)<br />
4.1.2.2 Stereographic processing<br />
If its geometry is completely unknown, a single image restitution of a 3D<br />
object is impossible. In this case the use of at least two images is necessary.<br />
According to the stereographic principle a pair of ‘stereo images’ can<br />
be viewed together, which produces a spatial (stereoscopic) impression of<br />
the object. This effect can be used to achieve a 3D restitution of, for<br />
example, facades. (See Figure 4.1.3.)<br />
Using ‘stereo pairs of images’ arbitrary shapes of a 3D geometry can be<br />
reconstructed as long as the area of interest is shown on both images. The<br />
camera directions should be almost parallel to each other to have a good<br />
stereoscopic viewing.<br />
Metric cameras with well known and calibrated interior orientation and<br />
negligible lens distortion are commonly used in this approach. To guarantee<br />
good results the ratio of stereo base (distance between camera<br />
positions) to the camera distance to the object should lie between 1:5 and<br />
1:15.<br />
Results of stereographic restitution can be:<br />
• 2D-plans of single facades;<br />
• 3D-wireframe and surface models;<br />
• lists of coordinates;<br />
• eventually complemented by their topology (lines, surfaces, etc.).<br />
Object distance<br />
Figure 4.1.3<br />
Stereo base
304 Pierre Grussenmeyer et al.<br />
Figure 4.1.4 Stereopair from CIPA-Testfield ‘Otto Wagner Pavillion Karlsplatz,<br />
Vienna’.<br />
Figure 4.1.5 2D façade plan derived from above stereo pair of images.<br />
4.1.2.3 Bundle restitution<br />
In many cases the use of one single stereo pair will not suffice to reconstruct<br />
a complex building. Therefore a larger number of photos will be<br />
used to cover an object as a whole. To achieve a homogeneous solution<br />
for the entire building and also to contribute additional measurements, a<br />
simultaneous solution of all the photo’s orientation is necessary.<br />
Another advantage is the possibility to perform an on-the-job calibration<br />
of the camera. This helps to increase the accuracy when using images<br />
of an unknown or uncalibrated camera. So this approach is no longer<br />
restricted to metric or even calibrated cameras, which makes the applica-
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
1<br />
2<br />
4<br />
5<br />
Architectural photogrammetry 305<br />
Bundle<br />
3 Convergent bundle<br />
1 2 3 4<br />
Figure 4.1.6 Examples of different configurations for bundle solution.<br />
tion of photogrammetric techniques much more flexible. It is also adjustable<br />
concerning the geometry of camera positions, meaning one is not<br />
forced to look for parallel views and stereo pair configuration. Convergent,<br />
horizontally, vertically or obliquely photos are now certainly suitable. Combination<br />
of different cameras or lenses can easily be done.<br />
The strategy of taking photos is that each point to be determined should<br />
be intersected by at least two rays of satisfactory intersection angle. This<br />
angle depends only upon the accuracy requirements. Additional knowledge<br />
of, for example, parallelism of lines, flatness of surfaces and rectangularity<br />
of features in space can be introduced in this process and helps to build<br />
a robust and homogeneous solution for the geometry of the object.<br />
The entire number of measurements and the full range of unknown parameters<br />
are computed within a statistical least squares adjustment. Due to<br />
the high redundancy of such a system it is also possible to detect blunders<br />
and gross errors, so not only accuracy but also reliability of the result<br />
will usually be increased.<br />
Bundle adjustment is a widespread technique in the digital architectural<br />
photogrammetry of today. It combines the application of semi-metric<br />
or even non-metric (amateur) cameras, convergent photos and flexible<br />
measurements in a common computer environment. Because of the adjustment<br />
process, the results are more reliable and accurate and very often<br />
readily prepared for further use in CAD environments.<br />
Results of bundle restitution are usually 3D-wireframe and surface<br />
models of the object or lists of coordinates of the measured points and their<br />
topology (lines, surfaces, etc.) for use in CAD and information systems.<br />
Visualizations and animations or so-called ‘photo-models’ (textured 3Dmodels)<br />
are also common results. Usually the entire object is reconstructed<br />
in one step and the texture for the surface is available from original photos<br />
(see §4.1.6).<br />
5<br />
7<br />
6
Figure 4.1.7 Examples of different images, different cameras, different lenses (from project Ottoburg, Innsbruck) to combine<br />
within a bundle solution (Hanke and Ebrahim, 1999).
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
4.1.3 Image acquisition systems<br />
Architectural photogrammetry 307<br />
Figure 4.1.8 Wireframe model and surface model as results of bundle restitution.<br />
4.1.3.1 General remarks<br />
<strong>Digital</strong> image data may be acquired directly by a digital sensor, such as a<br />
CCD array camera (see §1.5), for architectural photogrammetric work.<br />
Alternatively, it may be acquired originally from a photograph and then<br />
scanned (see §1.8).<br />
For the applications in architectural photogrammetry the use of cameras<br />
was for a long time determined by the use of expensive and specialized<br />
equipment (i.e. metric cameras). Depending on the restrictions due to the<br />
photogrammetric reconstruction process in former times, only metric<br />
cameras with known and constant parameters of interior orientation could<br />
be used. Their images had to fulfil special cases for the image acquisition<br />
(e.g. stereo normal case).<br />
Nowadays more and more image acquisition systems based on digital<br />
sensors are developed and available at reasonable prices on the market.<br />
The main advantage of these camera systems is the possibility to acquire<br />
digital images and directly process them on a computer.<br />
Figure 4.1.9 gives a principal overview of the basic photogrammetric<br />
systems for image acquisition and image processing in architectural photogrammetry.<br />
The classic, photographic cameras have their advantages in<br />
the unsurpassed quality of the film material and resolution and in the wellknown<br />
acquisition technique. The process of analytical photogrammetry<br />
makes a benefit of the knowledge and rich experiences of the human operator.<br />
On the other hand the pure digital data flow has not yet image
308 Pierre Grussenmeyer et al.<br />
Film-based<br />
camera<br />
Film<br />
development<br />
Analogue/Analytical<br />
<strong>Photogrammetry</strong><br />
Scanning /<br />
Digitisation<br />
<strong>Digital</strong> camera<br />
<strong>Digital</strong><br />
<strong>Photogrammetry</strong><br />
Figure 4.1.9 Image acquisition and image processing systems in architectural<br />
photogrammetry.<br />
acquisition devices comparable to film-based cameras. But this procedure<br />
allows a productive processing of the data due to the potential of automation<br />
and the simultaneous use of images and graphics. Furthermore, it<br />
allows a closed and therefore fast and consistent flow of data from the<br />
image acquisition to the presentation of the results. In addition, with<br />
the digitization of film a solution is available that allows the benefits<br />
of the high-resolution film to be merged with the benefits of the digital<br />
image processing. But the additional use of time for the separate process<br />
of digitization and the loss of quality during the scanning process are disadvantages.<br />
In the following sections the main photographic and digital image<br />
acquisition systems used in architectural photogrammetry are explained,<br />
examples are shown and criteria for their use are given.<br />
4.1.3.2 Photographic cameras<br />
From a photogrammetric point of view, film-based cameras can be subdivided<br />
into three main categories: metric cameras, stereo cameras and<br />
semi-metric cameras (see Table 4.1.1).<br />
Terrestrial metric cameras are characterized by a consequent opticalmechanical<br />
realization of the interior orientation, which is stable over a<br />
longer period. The image coordinate system is, like in aerial metric cameras,<br />
realized by fiducial marks. In architectural photogrammetry such cameras
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 309<br />
Table 4.1.1 Examples from the variety of film-based image acquisition systems<br />
Manufacturer Type Image format Lenses<br />
[mm 2 ] [mm]<br />
Metric cameras<br />
Hasselblad MK70 60 × 60 60, 100<br />
Wild P32 65 × 90 64<br />
Wild P31 100 × 130 45, 100, 200<br />
Zeiss UMK 1318 130 × 180 64, 100, 200, 300<br />
Stereo cameras<br />
Wild C 40/120 65 × 90 64<br />
Zeiss SMK 40/120 90 × 120 60<br />
Semi-metric cameras<br />
Rollei 3003 24 × 36 15–1000<br />
Leica R5 24 × 36 18–135<br />
Rollei 6006 60 × 60 40–350<br />
Hasselblad IDAC 55 × 55 38, 60, 100<br />
Pentax PAMS 645 40 × 50 35–200<br />
Linhof Metrica 45 105 × 127 90, 150<br />
Rollei R_metrica 102 × 126 75, 150<br />
Rollei LFC 230 × 230 165, 210, 300<br />
Geodetic Services CRC-1 230 × 230 120, 240, 450<br />
are less and less used. Amongst those cameras, which are still in practical<br />
use, are, for example, the metric cameras Wild P31, P32 and Zeiss UMK<br />
1318. Such image acquisition systems ensure a high optical and geometrical<br />
quality, but are also associated with high prices for the cameras<br />
themselves. In addition they are quite demanding regarding the practical<br />
handling. Besides the single metric cameras in heritage documentation,<br />
often stereo cameras are used. These cameras are composed of two calibrated<br />
metric cameras, which are mounted on a fixed basis in standard<br />
normal case. A detailed overview of terrestrial metric cameras can be found<br />
in many photogrammetric textbooks (e.g. Kraus, 1993).<br />
With the use of semi-metric cameras the employment of réseau techniques<br />
in photographic cameras was established in the everyday work of<br />
architectural photogrammetry. The réseau, a grid of calibrated reference<br />
marks projected on to the film at exposure, allows the mathematical<br />
compensation of film deformations, which occur during the process of<br />
image acquisition, developing and processing. Different manufacturers<br />
offer semi-metric cameras at very different film formats. Based on small<br />
and medium format SLR-cameras, systems exists from, for example, Rollei,<br />
Leica and Hasselblad. Their professional handling and the wide variety<br />
of lenses and accessories allow a fast and economic working on the<br />
spot. Semi-metric cameras with medium format offer a good compromise
310 Pierre Grussenmeyer et al.<br />
between a large-image format and established camera technique. An<br />
overview on semi-metric cameras is given in Wester-Ebbinghaus (1989).<br />
Often in architectural applications so called amateur cameras are used.<br />
This is not for dedicated photogrammetric projects, but in emergency cases,<br />
where no other recording medium was available or in case of destroyed<br />
or damaged buildings when only such imagery is available. Examples are<br />
given in, amongst others, Gruen (1976), Dallas et al. (1995) and Ioannidis<br />
et al. (1996). Due to the ongoing destruction of the world cultural heritage<br />
it will also be necessary in the future to reconstruct objects taken with<br />
amateur cameras (Waldheusl and Brunner, 1988).<br />
4.1.3.3 Scanners<br />
The digitization of photographic images offers a means to combine the<br />
advantages of film-based image acquisition (large image format, geometric<br />
and radiometric quality, established camera technique) with the advantages<br />
of digital image processing (archiving, semi-automatic and automatic<br />
measurement techniques, combination of raster and vector data).<br />
Scanners for the digitization of film material can be distinguished<br />
regarding different criteria. For example, regarding the type of sensor,<br />
either point, line or area sensor, or regarding the arrangement with respect<br />
to the scanned object as flatbed or drum scanner (see §1.8).<br />
For the practical use of scanners in architectural photogrammetric applications<br />
the problem of necessary and adequate scan resolution has to be<br />
faced. On the one hand the recognition of details has to be ensured and<br />
on the other hand the storage medium is not unlimited. This holds especially<br />
for larger projects. To scan a photographic film with a resolution<br />
equivalent to the film a scan resolution of about 12 m (2,100 dpi) is<br />
required. Thus, a scanned image from a medium format film (6 × 6 cm2 )<br />
has about 5,000 × 5,000 pixels. To hold this data on disk requires approximately<br />
25 Mbytes for a black-and-white scan and 75 Mbytes for a coloured<br />
image. For a scanned colour aerial image one would get a digital image<br />
of 20,000 × 20,000 pixels or 1.2 Gbytes. Even with the constant increasing<br />
size and decreasing costs for computer storage medium, this is a not to<br />
underestimate this factor in the planning of a project.<br />
For the use in architectural photogrammetry typically two different types<br />
of scanners are used, high-resolution photogrammetric scanners and<br />
desktop publishing scanners.<br />
The photogrammetric scanners are typically flatbed scanners, which have<br />
a high geometric resolution (5–12.5 m) and a high geometric accuracy<br />
(2–5 m). Currently there are just a few systems commercially available,<br />
which are offered mainly by photogrammetric companies. An overview on<br />
existing systems is given in Baltsavias and Bill (1994).<br />
The desktop publishing scanners (DTP) are not developed for photogrammetric<br />
use, but they are widely available on the market at low cost and
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 311<br />
they are developed and improved in a short time interval. DTP scanners<br />
have typically a scan size capability of A4 or A3 formats with a scan resolution<br />
of 300–1,200 dpi. The geometric resolution of these systems is about<br />
50 m. Despite this technical reduction compared to photogrammetric<br />
scanners, these scanners, which are low cost and easy to handle, can be<br />
used for photogrammetric purposes. This holds especially for calibrated<br />
systems, where geometric accuracy in the order of 5–10 m is feasible<br />
(Baltsavias and Waegli, 1996).<br />
Another possibility for the digitization and storage of film material is<br />
offered by the Photo-CD system. Small and medium format film can be<br />
digitized in a special laboratory and stored on CD-ROM. The advantage<br />
of such a system is the inexpensive and easy digitization and convenient<br />
data archiving. On the other hand the scanning process cannot be controlled<br />
or influenced and the image corners are usually not scanned. Thus<br />
the interior orientation of an image is nearly impossible to reconstruct.<br />
Investigations about the practical use of the Photo-CD system for digital<br />
photogrammetry are performed by Hanke (1994) and Thomas et al.<br />
(1995).<br />
4.1.3.4 CCD cameras<br />
The development of digital image acquisition systems is closely connected<br />
to the development of CCD sensors. The direct acquisition of digital images<br />
with a CCD sensor holds a number of advantages, which makes them<br />
interesting for photogrammetric applications. For example:<br />
• direct data flow with the potential of online processing;<br />
• high potential for automation;<br />
• good geometric characteristics;<br />
• independent of the film development process;<br />
• direct quality control of the acquired images;<br />
• low-cost system components.<br />
For photogrammetric applications mainly area-based CCD sensors are<br />
used. These sensors are produced for the commercial or industrial video<br />
market. Area-based CCD sensors are used in video cameras as well as in<br />
high resolution digital cameras for single exposures (still video cameras).<br />
Furthermore, there are specialized systems that use a scanning process for<br />
the image acquisition.<br />
Standard CCD video cameras have a number of 400–580 × 500–800<br />
sensor elements. With a pixel size of 7–20 m these cameras have an<br />
image format between 4.4 × 6.0 mm and 6.6 × 8.8 mm. Such cameras<br />
deliver a standardized, analogue video signal with 25–30 images per second.<br />
This signal can be displayed on any video monitor. For the photogrammetric<br />
processing of this signal, it has to be digitized by a frame grabber.
312 Pierre Grussenmeyer et al.<br />
Beside the entertainment industry such video cameras are mainly used for<br />
simple measurement or supervision tasks. The variety of systems on the<br />
market is enormous and can hardly be followed. Such systems are of importance<br />
for the documentation of destroyed objects of the world cultural<br />
heritage. Under certain circumstances the reconstruction of a building from<br />
video imagery is possible (Streilein, 1995).<br />
One of the newer developments of digital image acquisition devices are<br />
CCD video cameras with a digital output. These cameras offer an A/D<br />
conversion in the camera body and have storage capacity for one acquisition,<br />
so that it is possible to store the image. The digital image can be<br />
delivered to a computer, the transfer of the image data is practically free<br />
of noise.<br />
Due to the restrictions regarding resolution of CCD video cameras<br />
and due to the ongoing improvements in the CCD market, today more<br />
and more high-resolution digital cameras are used. Such cameras can be<br />
described as a combination of a traditional small-format SLR camera with<br />
a high-resolution CCD sensor replacing the film. The digital image data<br />
is stored directly in the camera body. In the photogrammetric community<br />
very well-known representatives of this type of camera are distributed from<br />
Kodak/Nikon under the Name DCS 420 and 460. They offer resolutions<br />
of 1,524 × 1,012 pixel and 3,060 × 2,036 pixel respectively. In addition,<br />
various manufacturers (e.g. Kodak, Leaf, and Rollei) offer camera systems<br />
with a resolution of about 2,000 × 2,000 pixels. The main advantage of<br />
such systems is the fast and easy image acquisition. This is achieved due<br />
to the fact that image acquisition, A/D conversion, storage medium and<br />
power supply are combined in one camera body. This allows one to transfer<br />
the images immediately to a computer and to judge the quality of the<br />
acquired images or directly process them.<br />
Scanning cameras improve the resolution of the final image by sequential<br />
scanning procedure. Two scanning principles are distinguished, micro<br />
scanning and macro scanning (Lenz and Lenz, 1993). Micro scanning<br />
cameras use the principle of moving a CCD area sensor in parts of the<br />
pixel size over the sensor. The final image will be calculated from the<br />
different single images taken after the movement. The format of the final<br />
image is the same as that of the single images. The resolution of the final<br />
image is increased in horizontal and vertical directions by the number of<br />
the single images taken in each direction (microscan factor). Macro scanning<br />
cameras on the other hand move a CCD area sensor over a large<br />
image format. The position of the single images taken is determined either<br />
optical-numerically or mechanically, such that the single images can be<br />
combined into one high-resolution image. Both methods have in common<br />
that they should be used when object and camera do not move during the<br />
period of image acquisition (typically several seconds). Furthermore, the<br />
illumination conditions should remain constant during that period. This<br />
is practically achievable only under laboratory conditions, so that such
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 313<br />
Table 4.1.2 Examples from the spectrum of digital image acquisition devices<br />
Manufacturer Type Number of pixel Image format<br />
(HV) [mm 2 ]<br />
Pulnix TM-560 582 × 500 8.8 × 6.6<br />
Canon ION RC 560 795 × 576 8.8 × 6.6<br />
Sony XC-77 768 × 593 8.8 × 6.6<br />
JVC GRE-77 728 × 568 6.4 × 4.8<br />
Canon EOS-DCS 3 1268 × 1012 20.5 × 16.5<br />
Kodak Megaplus-1.4 1320 × 1035 9.0 × 7.0<br />
Agfa ActionCam 1528 × 1148 16.5 × 12.4<br />
Kodak DCS 420 1524 × 1012 14.0 × 9.0<br />
Kodak Megaplus-4.2 2048 × 2048 18.4 × 18.4<br />
Rollei ChipPack 2048 × 2048 31.0 × 31.0<br />
Kontron ProgRes 3000 2994 × 2320 8.6 × 6.5<br />
Kodak DCS 460 3060 × 2036 30.0 × 20.0<br />
Canon EOS-DCS 1 3060 × 2036 27.6 × 18.4<br />
Dicomed BigShot 4096 × 4096 60.0 × 60.0<br />
Kodak Megaplus-16.8 4096 × 4096 36.8 × 36.8<br />
Agfa StudioCam 4500 × 3648 36.0 × 29.0<br />
RJM JenScan 4500 × 3400 8.8 × 6.6<br />
Jenoptik Eyelike 6144 × 6144 28.6 × 28.6<br />
Rollei RSC 7000 × 5500 55.0 × 55.0<br />
Zeiss UMK HighScan 15414 × 11040 166.0 × 120.0<br />
systems have minor importance for applications in architectural photogrammetry.<br />
For applications in architectural photogrammetry line-based sensors can<br />
also be used for the generation of high-resolution images. An example of<br />
this type of sensor used for architectural photogrammetry is given in Reulke<br />
and Scheele (1997). In this example several buildings are recorded using<br />
the WAOSS/WAAC CCD line scanner, which incorporates three line<br />
sensors of 5,184 pixels each.<br />
In Table 4.1.2 a few examples from the great variety of digital image<br />
acquisition systems on the market, which are suitable for applications<br />
in digital architectural photogrammetry, are given. This compilation is<br />
naturally incomplete and is just mentioned to give an idea about the<br />
variety. A good overview on digital image acquisition systems is nowadays<br />
available on different Internet pages, e.g. Plug-In Systems (2000),<br />
PCWebopaedia (2000), or the Internet pages of the different manufacturers.<br />
4.1.3.5 Which camera to use?<br />
For the question which camera to use for a specific photogrammetric task<br />
for heritage documentation, there is basically no common answer or simple<br />
rule. Often a photogrammetric project is as complex as the object itself
314 Pierre Grussenmeyer et al.<br />
and more often the use of an image acquisition device is determined by<br />
the budget of the project (read: use a camera system that is already available<br />
at no cost).<br />
However, for a successful photogrammetric project several aspects have<br />
to be taken into account. For example, the maximum time limit for the<br />
image acquisition on the spot and for the (photogrammetric) processing<br />
of the data afterwards. Further criteria can be: the need for colour images<br />
or black-and-white images, the requested accuracy of the final model, the<br />
smallest object detail that can be modelled, the minimum and maximum<br />
of images for the project, the mobility and flexibility of the image acquisition<br />
system or the integration into the entire production process. But<br />
after all the price of the image acquisition system and the possibilities<br />
for further processing of the image data remain as the major factors for<br />
selecting a specific image acquisition system for a specific project.<br />
4.1.4 Overview of existing methods and systems for<br />
architectural photogrammetry<br />
4.1.4.1 General remarks<br />
Architectural photogrammetry and aerial photogrammetry don’t have the<br />
same applications and requirements. Most of the commercial digital<br />
photogrammetric workstations (DPWs) are mainly dedicated to stereoscopic<br />
image measurement, aerial triangulation, the digital terrain model<br />
(DTM) and orthoimages production from aerial and vertical stereopair<br />
images. A survey of these DPWs is presented by Plugers (1999). In this<br />
section we consider systems and methods which are rather low cost<br />
comparing to those well-known products developed mainly for digital<br />
mapping. Software packages for architectural photogrammetry may use<br />
different image types, obtained directly by CCD cameras or by scanning<br />
small and medium format metric or non-metric slides (see §4.1.3). The<br />
quality of digital images directly influences the final result: the use of lowresolution<br />
digital cameras or low-priced scanners may be sufficient for<br />
digital 3D visual models but not for a metric documentation. The systems<br />
may be used by photogrammetrists as well as by architects or other specialists<br />
in historic monument conservation, and run on simple PC-systems that<br />
suffice for special tasks in architectural photogrammetry.<br />
According to the specific needs in architectural documentation, the<br />
different kinds of systems are based either on digital image rectification,<br />
or on monoscopic multi-image measurement or on stereoscopic image<br />
measurement (Fellbaum, 1992). Software of such systems is advanced in<br />
such a way that mass restitution and modelling is possible, if digital images<br />
are provided in a well-arranged way. Only some software packages are<br />
mentioned in this paragraph, and the aim is not to make a survey of<br />
existing systems.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
To compare different systems, the following topics can be considered<br />
(CIPA, 1999):<br />
• the handling of a system;<br />
• the flow of data;<br />
• the management of the project;<br />
• the import and export of data (image formats, parameter of interior<br />
and exterior orientation, control information, CAD information);<br />
• the interior orientation;<br />
• the exterior orientation (one step or two steps);<br />
• the reconstruction of the object;<br />
• the derived results in terms of topology, consistency, accuracy and reliability;<br />
• the amount of photogrammetric knowledge necessary to handle the<br />
system.<br />
4.1.4.2 Recommendation for simple photogrammetric<br />
architectural survey<br />
For simple photogrammetric documentation of architecture, simple rules<br />
that are to be observed for photography with non-metric cameras have<br />
been written, tested and published by Waldheusl and Ogleby (1994).<br />
These so-called ‘3 × 3 rules’ are structured in:<br />
1 three geometrical rules:<br />
i preparation of control information;<br />
ii multiple photographic all-around coverage;<br />
iii taking stereopartners for stereo-restitution.<br />
Architectural photogrammetry 315<br />
Figure 4.1.10 Ground plan of a stable bundle block arrangement all around<br />
building, as recommended in the 3 × 3 rules<br />
[see http://www.cipa.uibk.ac.at].
316 Pierre Grussenmeyer et al.<br />
2 three photographic rules:<br />
i the inner geometry of the camera has to be kept constant;<br />
ii select homogeneous illumination;<br />
iii select most stable and largest format camera available.<br />
3 three organizational rules:<br />
i make proper sketches;<br />
ii write proper protocols;<br />
iii don’t forget the final check.<br />
Usually, metric cameras are placed on a tripod, but shots with small or<br />
medium format equipment are often taken ‘by hand’. Recently, digital<br />
phototheodolites combining total-station and digital cameras have been<br />
developed by Agnard et al. (1998), and Kasser (1998). <strong>Digital</strong> images are<br />
then referenced from object points or targets placed in the field. In this<br />
way, the determination of the exterior orientation is simple and the images<br />
are directly usable for restitution.<br />
4.1.4.3 <strong>Digital</strong> image rectification<br />
Many parts of architectural objects can be considered as plane. In this<br />
case, even if the photo is tilted with regard to the considered plane of the<br />
object, a unique perspective is enough to compute a rectified scaled image.<br />
We need at least four control points defined by their coordinates or<br />
distances in the object plane (§2.3.2).<br />
The Rolleimetric MSR software package [http://www.rolleimetric.de]<br />
provides scale representations of existing objects on the basis of rectified<br />
digital images (Figures 4.1.11 (a), (b) and (c)). The base data is usually<br />
one or more photogrammetric images and/or amateur photographs of the<br />
object that are rectified at any planes defined by the user. Simple drawings<br />
(in vector-mode), image plans (in raster-mode) are processed as a<br />
result of the rectification.<br />
Basically, the major stages encountered in the rectification of photography<br />
are as follows (Bryan et al., 1999):<br />
• site work (photography and control);<br />
• scanning;<br />
• rectification;<br />
• mosaicking;<br />
• retouching;<br />
• output;<br />
• archive storage.<br />
Photographs of building façades should be taken the most perpendicular<br />
to the reference planes and only the central part of the image should<br />
be considered for a better accuracy.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
(a)<br />
(b)<br />
Architectural photogrammetry 317<br />
Figure 4.1.11 (a) Selection of a ‘plane’ part of the object. (b) Control distances<br />
for the plane definition. (c) Rectified image as an extract of (a).<br />
Commercial CAD/CAM software packages often include image handling<br />
tools and also allow simple image transformation and rectification. But<br />
they seldom consider camera distortions, as opposed to photogrammetric<br />
software.<br />
In the case of a perspective rectification, radial image displacements in<br />
the computed image will occur for points outside the reference. The rectification<br />
obviously fails if the object isn’t somewhat plane.<br />
Some packages include functions for the photogrammetric determination<br />
of planes according to the multi-image process (§4.1.4.4) from two<br />
or three photographs that capture an object range from different viewpoints.<br />
<strong>Digital</strong> image maps can be produced by assuming the object surface<br />
and photo rectification. In the resulting orthophoto, the object model is<br />
represented by a digital terrain model (Pomaska, 1998). Image data of<br />
different planes can be combined into digital 3D computer models for visualization<br />
and animation with the help of photo editing or CAD software.<br />
ELSP from PMS [http://www.pms.co.at] and Photoplan [http://www.<br />
photoplan.net] are other examples of commercial systems particularly dedicated<br />
to rectification.<br />
Besides digital image rectification and orthoimaging, single images for<br />
3D surfaces of known analytical expression may lead to products in raster<br />
form, based on cartographic projections. Raster projections and developments<br />
of non-metric imagery of paintings on cylindrical arches of varying<br />
diameters and spherical surfaces are presented in Karras et al. (1997) and<br />
Egels (1998).<br />
4.1.4.4 Monoscopic multi-image measurement systems<br />
Photogrammetric multi-image systems are designed to handle two or more<br />
overlapping photographs taken from different angles of an object (see<br />
§4.1.2.3). In the past, these systems were used with analogue images<br />
(c)
318 Pierre Grussenmeyer et al.<br />
enlarged and placed on digitizing tablets. Presently, the software usually<br />
processes image data from digital and analogue imaging sources (réseau,<br />
semi-metric or non-metric cameras). Scanners are used to digitize the<br />
analogue pictures. Film and lens distorsions are required to perform metric<br />
documentation. Monoscopic measurements are achieved separately on each<br />
image. These systems don’t give the opportunity of conventional stereophotogrammetry.<br />
For the point measurements and acquisition of the object<br />
geometry, systems can propose support:<br />
• for the automatic réseau cross-measurement;<br />
• for the measurement of homologous points through the representation<br />
of line information in the photogrammetric images and epipolar<br />
geometry.<br />
The acquisition of linear objects can be directly evaluated due to superimposition<br />
in the current photogrammetric image. These systems are<br />
designed for multi-image bundle triangulation (including generally simultaneous<br />
bundle adjustment with self-calibration of the used cameras, see<br />
§2.3). The main differences between the systems consist in the capabilities<br />
of the calculation module to combine additional parameters and<br />
knowledge, like directions, distances and surfaces. The main contribution<br />
of digital images in architectural photogrammetry is the handling of<br />
textures. The raster files are transformed into object surfaces and digital<br />
image data is projected on to a three-dimensional object model. Some<br />
systems are combined with a module of digital orthorectification.<br />
The Canadian Photomodeler Software Package developed by Eos Systems<br />
is well known as a low-cost 3D-measurement tool for architectural and<br />
archeological applications [http://www.photomodeler.com]. Photomodeler<br />
(Figure 4.1.12) is a Windows-based software that allows measurements<br />
and transforms photographs into 3D models. The basic steps in a project<br />
performed with Photomodeler are:<br />
• shoot two or more overlapping photographs from different angles of<br />
an object;<br />
• scan the images into digital form and load them into Photomodeler;<br />
• using the point and line tools, mark on the photographs the features<br />
you want in the final 3D model;<br />
• reference the points by indicating which points on different photographs<br />
represent the same location on the object;<br />
• process referenced data (and possibly the camera) to produce 3D<br />
model;<br />
• view the resulting 3D model in the 3D viewer;<br />
• extract coordinates, distances and areas measurements within Photomodeler;<br />
• export the 3D model to rendering, animation or CAD program.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 319<br />
Figure 4.1.12 Photomodeler’s measuring module (project ‘Ottoburg, Innsbruck’,<br />
see §4.1.6).<br />
Fast camera calibration based on a printable plane pattern can be set up<br />
separately to compute the camera’s focal length, principal point, digitizing<br />
aspect ratio and lens distortion. Images from standard 35 mm film cameras<br />
digitized with Kodak Photo-CD, negative scanner or flatbed scanner as<br />
well as from digital and video cameras can be used in Photomodeler. The<br />
achieved accuracy obtained by Hanke and Ebrahim (1997) for distances<br />
between points lies in the range of 1:1,700 (for a 35 mm small format<br />
‘amateur’ camera, without lens distortion compensation) to 1:8,000 (for<br />
a Wild P32 metric camera) and shows promising results.<br />
Examples of systems based on the same concept are:<br />
• KODAK <strong>Digital</strong> Science Dimension Software [http://www.kodak.com],<br />
with single and multiple image capabilities.<br />
• 3D BUILDER PRO [http://aay.com/release.htm], with a ‘constraintbased’<br />
modelling software package.<br />
• SHAPECAPTURE [http://www.shapequest.com] offers target and<br />
feature extraction, target and feature 3D coordinate measurement, full<br />
camera calibration, stereo matching and 3D modelling.<br />
• CANOMA [http://www.metacreations.com] from Meta Creations is a<br />
software intended for creating photorealistic 3D models from illustrations<br />
(historical materials, artwork, hand-drawn sketches, etc.),<br />
scanned or digital photographs. Based on an image-assisted technique,<br />
it enables one to attach 3D wireframe primitives and to render a 3D<br />
image by wrapping the 2D surfaces around these primitives.
320 Pierre Grussenmeyer et al.<br />
Some other systems, also based on monoscopic multi-image measurement,<br />
are not mainly dedicated to the production of photomodels. In<br />
general, they use CAD models for the representation of the photogrammetric-generated<br />
results. For examples there are:<br />
• CDW from Rolleimetric [http://www.rolleimetric.de] which is mainly<br />
a measuring system and doesn’t handle textures. Data are exported to<br />
the user’s CAAD system by interfaces. MSR 3D , also proposed by Rolleimetric,<br />
is an extension of MSR and is based on a CDW multi-image<br />
process (two or three photographs) for the determination of the different<br />
object-planes and the corresponding rectified images.<br />
• Elcovision12 from PMS [http://www.pms.co.at] can run standalone or<br />
be directly interfaced with any CAAD application.<br />
• PICTRAN from Technet (Berlin, Germany) [www.technet-gmbh.com]<br />
which includes bundle block adjustment and 3D restitution as well as<br />
rectification and digital orthophoto.<br />
• PHIDIAS, proposed by PHOCAD (Aachen, Germany) is integrated in<br />
the Microstation CAD package.<br />
• ORPHEUS (ORIENT-<strong>Photogrammetry</strong> Engineering Utilities System),<br />
proposed by the Institute of <strong>Photogrammetry</strong> and Remote Sensing of<br />
the Vienna University of Technology (Austria), is a digital measurement<br />
module running with the ORIENT software, linked with the<br />
SCOP software for the generation of digital orthophotos. It allows<br />
one, more particularly, to handle large images by the creation of image<br />
pyramids.<br />
The prototype DIPAD (digital system for photogrammetry and architectural<br />
design) is based on the idea of using CAD models in a priori and<br />
a posteriori modes. Therefore CAD models are not only used for the<br />
graphical visualization of the derived results, but a complete integration<br />
between CAAD environment and photogrammetric measurement routines<br />
is realized. No manual measurements have to be performed during the whole<br />
analysis process. The system allows the three-dimensional reconstruction of<br />
objects or parts of it as an object-oriented approach in a CAAD-controlled<br />
environment. The task of object recognition and measurement is solved,<br />
with the image interpretation task performed by a human operator and the<br />
CAD-driven automatic measurement technique derives the precise geometry<br />
of the object from an arbitrary number of images. A human being uses<br />
intuitively, while looking at an image, his/her entire knowledge of the real<br />
world and is therefore able to select the information from or add missing<br />
information to the scene, which is necessary to solve the specific task.<br />
By choosing a combined top-down and bottom-up approach in feature<br />
extraction, a coarsely given CAD model of the object is iteratively refined<br />
until finally a detailed object description is generated. The semi-automatic<br />
HICOM method (human for interpretation and computer for measurement)
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
is able to simplify the analysis process in architectural photogrammetry<br />
without degradation in accuracy and reliability of the derived results.<br />
Typical problems for the automatic analysis of image data, which are very<br />
common in outdoor scenes, like occlusions of the object by itself or due to<br />
other objects (e.g. persons, vegetation), shadow edges and reflections, are<br />
detected with this method and attributed during further processing.<br />
Examples for the performance of this method are given in, for example,<br />
Streilein (1996) and (Streilein and Niederöst, 1998).<br />
4.1.4.5 Stereoscopic image measurement systems<br />
Architectural photogrammetry 321<br />
Figure 4.1.13 Screenshot of DIPA during the reconstruction process (interaction<br />
takes only place in the lower left CAD environment).<br />
From analytical to digital<br />
<strong>Digital</strong> stereoscopic measuring systems follow analytical stereoplotters well<br />
known as the more expensive systems. Many plottings are still done on<br />
analytical stereoplotters for metric documentation but as the performance<br />
and ability of digital systems increase and allow mass restitution. As<br />
textures are increasingly required for 3D models, digital photographs and<br />
systems are getting more and more important.
322 Pierre Grussenmeyer et al.<br />
Stereoscopy<br />
Systems presented in the former paragraph allow more than two images<br />
but homologous points are measured in monoscopic mode. Problems<br />
may occur for objects with less texture when no target is used to identify<br />
homologous points. Only stereo-viewing allows in this case a precise<br />
3D measurement. Therefore stereopairs of images (close to the normal<br />
case) are required. Systems can then be assimilated to 3D plotters for the<br />
measuring of spatial object coordinates. 3D measurements are required for<br />
the definition of digital surface models that are the basis of the orthophotos.<br />
Usually, proposed solutions for stereo-viewing devices are:<br />
• the split-screen configuration using a mirror stereoscope placed in front<br />
of the screen;<br />
• the anaglyph process;<br />
• the alternating of the two images on the full screen (which requires<br />
active glasses);<br />
• the alternating generation of the two synchronized images on a polarized<br />
screen (which requires polarized spectacles).<br />
Automation and correlation<br />
In digital photogrammetry, most of the measurements can be done automatically<br />
by correlation. The task is then to find the position of a geometric<br />
figure (called the reference matrix) in a digital image. If the approximate<br />
position of the measured point is known in the image, then we can define<br />
a so-called search matrix. Correlation computations are used to determine<br />
the required position in the digital image. By correlation in the subpixel<br />
range the accuracy of positioning is roughly one order of magnitude better<br />
than the pixel size. Usually, the correlation process is very efficient on<br />
architectural objects, due to textured objects. Correlation functions can be<br />
implemented in the different steps of the orientation:<br />
• fiducial marks or réseau crosses can be measured automatically in the<br />
inner orientation;<br />
• measurement of homologous points can be automated by the use of<br />
the correlation process both in the exterior orientation, and in the<br />
digital surface model and stereoplotting modules.<br />
The correlation function is a real step forward compared to manual<br />
measurements applied in analytical photogrammetry. The quality of the<br />
measurement is usually given by a correlation factor.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 323<br />
Model orientation<br />
Systems currently used for aerial photogrammetry run with stereopairs of<br />
images but their use isn’t always possible in architectural photogrammetry.<br />
In many cases, systems cannot handle different types and scales of images.<br />
The orientation module may fail either because photographs and control<br />
points are defined in a terrestrial case, or due to the non-respect of the<br />
approximate normal case. For the exterior orientation, some systems<br />
propose solutions based on typical two-step orientation (relative and<br />
absolute), as known in analytical photogrammetry. But bundle adjustment<br />
is increasingly proposed and allows an orientation in only one step. A direct<br />
linear transformation is sometimes used to compute approximate parameters<br />
of the orientation. Some systems propose automatic triangulation.<br />
But due to the configuration of the set of photographs in architectural<br />
photogrammetry, their use requires many manual measurements. After<br />
the relative orientation, often normalized images can be computed. Either<br />
original or normalized images can be used for further processes.<br />
Stereo-digitizing and data collection<br />
3D vectorization allows plottings (Figure 4.1.14) and digital surface model<br />
generation, with more or less semi-automatic procedures depending on the<br />
systems. Image superimposition is possible with almost every DPW. Some<br />
Figure 4.1.14 <strong>Digital</strong> stereoplotting with image superimposition.
324 Pierre Grussenmeyer et al.<br />
systems can be connected on-line to commercial CAD software packages<br />
and use their modelling procedures. Orthophotos can be generated from<br />
the different models but for architectural application, photomodels are<br />
usually carried out with multi-image systems (see Figures 4.1.6 and 4.1.7).<br />
Examples of low-cost PC-systems based on stereoscopic measurements<br />
with known applications in close-range architectural photogrammetry, but<br />
only handling nearly normal case stereopairs, are:<br />
• The digital video plotter (DVP) [http://www.dvp-gs.com]. It was one<br />
of the first systems proposed (Agnard et al., 1988). It is optimized for<br />
large-scale mapping in urban areas but architectural projects have been<br />
presented by DVP Geomatic Systems Inc.<br />
• Photomod from Raccurs (Moscow, Russia) has a high degree of automatization,<br />
compared to other systems. The system is known as DPW<br />
for aerial photogrammetry. Orthophotos of facades with Photomod<br />
have been presented by Continental Hightech Services [http://www.chscarto.fr].<br />
• Imagestation SSK stereo Softcopy Kit from Intergraph Corporation is<br />
proposed as a kit to convert a PC into a low-cost DPW. Different<br />
modules of Intergraph ImageStation are available.<br />
Several systems have been developed by universities during recent years.<br />
Some of them are:<br />
• VSD as video stereo digitizer (Jachimski, 1995) is well known for<br />
architectural applications; VSD is a digital autograph built on the basis<br />
of a PC. It is suitable for plotting vector maps on the basis of pairs<br />
of monochrome or colour digital images as well as for vectorization<br />
of orthophotographs. The stereoscopic observation is based on a simple<br />
mirror stereoscope.<br />
• POIVILLIERS ‘E’ developed by Yves Egels (IGN-ENSG, Paris, France)<br />
runs under DOS. Stereo-viewing is possible by active glasses connected<br />
to the parallel port of the PC or by anaglyphs. The system is very efficient<br />
for large images and high pointing accuracy is available due to<br />
a sub-pixel measuring module. Colour superimposition of plottings is<br />
also proposed. The system runs on aerial images as well as on terrestrial<br />
ones.<br />
• Mapping from stereopairs within Autocad R14 is proposed by Greek<br />
Universities (Glykos et al., 1999).<br />
• TIPHON is a Windows application developed at ENSAIS (Polytechnicum<br />
of Strasbourg, France) for two-image based photogrammetry<br />
(stereopair or convergent bundle) with different kinds of cameras<br />
(Grussenmeyer and Koehl, 1998). The measurements on the images<br />
are manual or semi-automatic by correlation. A stereoscope is used if<br />
stereoscopic observations are required.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
• ARPENTEUR as ‘ARchitectural PhotogrammEtry Network Tool for<br />
EdUcation and Research’ is a platform-independent system available<br />
on the web by a simple internet browser (Drap and Grussenmeyer,<br />
2000). The system is an adaptation of TIPHON to the Internet World<br />
and is particularly dedicated to architectural applications. ARPEN-<br />
TEUR is a web-based software package utilizing HTTP and FTP<br />
protocols. The photogrammetric adjustment and image processing<br />
routines are written in JAVA. Different solutions are available for<br />
the orientation of the digital stereopairs. The concept of running<br />
photogrammetric software on the Internet is extended by a new<br />
approach of architectural photogrammetry 3D modelling. The architectural<br />
survey is monitored by underlying geometrical models stemming<br />
from architectural knowledge. Image correlation, geometrical<br />
functions and photogrammetry data are combined to optimize the modelling<br />
process. The data of the stereoplotting are directly superimposed<br />
on the images and exported towards CAD software packages<br />
and VRML file format. ARPENTEUR is usable via the Internet at<br />
[http://www.arpenteur.net].<br />
4.1.5 3D object structures<br />
Architectural photogrammetry 325<br />
4.1.5.1 General remarks<br />
If a person is asked to describe an object, he/she solves the problem typically<br />
by describing all the single components of the object with all their<br />
attributes and properties and the relations they have with respect to each<br />
other and to the object. In principle computer representations and models<br />
are nothing other than the analogue description of the object, only the<br />
human language is replaced by mathematical methods. All kinds of representations<br />
describe only a restricted amount of attributes and each finite<br />
mathematical description of an object is incomplete.<br />
Data models are necessary in order to process and manipulate real-world<br />
objects with the computer. The data models are abstractions of realworld<br />
objects or phenomena. Abstractions are used in order to grasp or<br />
manipulate the complex and extensive reality. Each attempt to represent<br />
reality is already an abstraction. The only complete representation of a<br />
real-world object is the object itself. Models are structures, which combine<br />
abstractions and operands into a unit useful for analysis and manipulation.<br />
Using models the behaviour, appearance and various functions of an<br />
object or building can be easily represented and manipulated. Prerequisite<br />
for the origin of a model is the existence of an abstraction. Each model<br />
needs to fulfil a number of conventions to work with it effectively. The<br />
higher the degree of abstraction the more conventions have to be fulfilled.<br />
CAAD models represent in an ideal way the building in form, behaviour<br />
and function as a logical and manipulable organism.
326 Pierre Grussenmeyer et al.<br />
The data of the computer internal representation, which is sorted according<br />
to a specific order (‘data structure’), forms the basis for software<br />
applications. The data basis is not directly accessed, but via available model<br />
algorithms, which allow the performance of complex functions by transforming<br />
them into simple basic functions according to a defined algorithm.<br />
The representation of a real-world object in a computer-oriented model is<br />
a synthesis of data structure and algorithms. Depending on extent and<br />
amount of information of the data, an object can be represented as a dataintensive<br />
or an algorithm-intensive model (Grätz, 1989). The most important<br />
role in the definition of models plays a proper balance between correctness<br />
and easy handling.<br />
4.1.5.2 Classification of 3D models<br />
In principle 3D models can be subdivided into three independent classes, the<br />
wireframe model, the surface model and the solid model (see Figure 4.1.15).<br />
The division is based on the different computer internal representation<br />
schemes and is therefore also for the application areas of these models.<br />
Wireframe models are defined by the vertices and the edges connecting<br />
these vertices. They fix the outlines of an object and allow a view through<br />
from any point of view. This is an advantage for simple objects, but reduces<br />
the readability of more complex objects. This representation is therefore<br />
often used for simple objects.<br />
Surface models represent the object as an ordered set of surfaces in threedimensional<br />
space. Surface models are mainly used for the generation of<br />
models, whose surfaces consist of analytical not easily describable faces<br />
having different curvatures in different directions. This is often the case<br />
for models of automobiles, ships or aeroplanes.<br />
Volumetric models represent three-dimensional objects by volumes. The<br />
data structure allows the use of Boolean operations as well as the calculation<br />
of volume, centre of gravity and surface area (Mäntylä, 1988).<br />
Surface modelling is the most demanding but also the most calculationintensive<br />
way of modelling. Solid models always represent the hierarchy<br />
of the object, in which the primitives and operations are defined.<br />
3D model<br />
wireframe model surface model volume model<br />
Figure 4.1.15 Overview of 3D models.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 327<br />
Each of the classes mentioned above has its specific advantages and<br />
disadvantages. Depending on the task the advantages and disadvantages<br />
are more or less important. Therefore it is not possible to make a general<br />
statement, which of the classes is the best representation of a real-world<br />
object.<br />
Every representation of an object is a more or less exact approximation<br />
of the reality. A true-reality representation of a building considers all important<br />
attributes of the design. Looking at the section of a box object (see<br />
Figure 4.1.16) shows the differences of the different representations.<br />
• The wireframe model represents the box as a quantity of vertices and<br />
edges. The section shows a quantity of non-connected points. This<br />
representation is reality-true if one is interested in the general form or<br />
in the position of the box.<br />
• The surface model describes the box as a combination of vertices,<br />
edges and surfaces. The section shows a quantity of points and lines.<br />
This representation is reality-true if one is interested in the appearance<br />
of the surfaces.<br />
• The volumetric model shows the box as a quantity of vertices, edges,<br />
faces and volume elements. The section shows points, lines and faces.<br />
This representation is reality-true if one is interested in mass properties,<br />
dynamic properties or material properties. For this purpose<br />
additional information that does not belong to the pure geometry of<br />
the object has to be evaluated.<br />
Figure 4.1.16 Reality-true representation of a box as wireframe, surface and<br />
volumetric model.
328 Pierre Grussenmeyer et al.<br />
Figure 4.1.17 Example for ambiguity of a wireframe representation.<br />
4.1.5.3 Wireframe models<br />
The simplest forms of three-dimensional models are wireframe models. The<br />
computer internal model is confined basically on the edges of the object.<br />
Between those edges exists no relation, a relation to faces is not defined.<br />
Information about the inside and the outside of the object is not available.<br />
Points and edges are the only geometric elements of a wireframe model and<br />
are represented in the computer by list structures.<br />
If a quantity of edges is stored with the condition that these edges should<br />
form the edges of a real-world object, it is often not unequivocal which<br />
object is represented. There exists more than one object that have the same<br />
edges as elements (see Figure 4.1.17).<br />
The wireframe model allows therefore only an incomplete and ambiguous<br />
representation of a real-world object, which requires human interpretation.<br />
Manipulations are only possible on the basis of object edges (not<br />
object oriented) with the result that, as a result of geometric operations,<br />
nonsensical objects can be modelled. For example, deleting an edge of a<br />
cube results in a wireframe model which has no representation in the real<br />
world.<br />
4.1.5.4 Surface models<br />
The main information of the data structure of surface models is carried<br />
in the single surfaces of the model. These surfaces are typically not easy<br />
to describe by analytical means. A vector representation is also not possible.<br />
As a mathematical representation of such surfaces, approximate procedures
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Z<br />
Y<br />
X<br />
Architectural photogrammetry 329<br />
v u<br />
x = x (u, v)<br />
y = y (u, v)<br />
z = z (u, v)<br />
Figure 4.1.18 Parametric representation of a surface (adapted from Grätz, 1989).<br />
are often used, where the surface is represented between given points by<br />
a parametric function. The three-dimensional coordinates of the surface<br />
are a function of surface coordinates, which are used to parameterize the<br />
surface (see Figure 4.1.18).<br />
Quite a number of mathematical procedures to handle surface models<br />
are available, such as Bézier approximations, spline and B-spline interpolations.<br />
The surfaces in this data structure are quasi-independent elements, which<br />
means that they have no relations to each other. Surfaces do not share<br />
edges, the edges belong to the geometric description of the surfaces. A<br />
relation between a surface and a volume does not exist. Therefore models<br />
can be constructed, which are composed of a number of surfaces, but have<br />
no representation in the real world.<br />
4.1.5.5 Volume models<br />
The third class of 3D models are the volume models, which covers quite<br />
a number of different classes, such as parametric representation, sweep<br />
representation, octrees, cell decomposition, boundary representations,<br />
constructive solid geometry and hybrid models.
330 Pierre Grussenmeyer et al.<br />
Characteristic for volume models is the computer internal volumetric<br />
representation of three-dimensional physical objects of the real word. These<br />
models are complete and non-ambiguous. Complete in this respect means<br />
that it is impossible to define or represent a volume with a missing surface<br />
or edge. On the other hand it is impossible to introduce surfaces or edges<br />
into the model, which do not belong to a volume. The produced computer<br />
model is in a technical sense a real representation of the reality. Therefore<br />
nonsense objects, like in wireframe and surface models, are impossible.<br />
Amongst the most commonly used volume models are the boundary<br />
representation (BRep) and the constructive solid geometry (CSG).<br />
The boundary representation (BRep) describes a three-dimensional object<br />
by its boundary elements (faces, edges and vertices). A boundary representation<br />
holds implicit a hierarchy in its data structure. An object is<br />
defined by its faces, the faces by its edges, the edges by its vertices and<br />
the vertices by three coordinates. A BRep model is organized in a topologic<br />
part and in a geometric part. Whereas the topologic part describes<br />
the neighbourhood relations between vertices, edges and faces, the<br />
geometric part holds the geometric information about the topologic<br />
elements (e.g. the three coordinates of the vertices). For the topologic relations<br />
have to obtain the following rules:<br />
• each object is in relation to a number of faces, which bound it;<br />
• each face is in relation to a number of edges, which bound it;<br />
• each edge separates two faces from each other;<br />
• each edge is in relation to a number of vertices, which limits it.<br />
The topologic description of the object describes it as an arbitrary<br />
deformable object, which solidifies according to the geometric description.<br />
Constructive solid geometry (CSG) describes the object instead of its<br />
bounding faces and edges by a set of defined volumetric primitives and<br />
basic operations of these primitives. The basic idea is that complex objects<br />
are represented as ordered additions, subtractions and intersections of<br />
simple objects. The primitives, which are used to sequentially build up the<br />
model, are directly visible in the data structure. Hence, basic volumes like<br />
cuboids, cones, globes, cylinders, etc. are contained in the data structure<br />
as primitives and directly accessible. They form the lowest level of geometric<br />
elements. An object is defined by the following grammar:<br />
:: |<br />
argument |<br />
<br />
:: cube | cuboid | cylinder | cone | globe | . . .<br />
:: translation | rotation | scaling<br />
:: | | .
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 331<br />
According to this grammar, each object is describable as a binary tree,<br />
where the leaves represent the primitives and the nodes represent the operations<br />
and transformations. The first node represents the modelled object.<br />
With this scheme the hierarchy of an object is implicitly given.<br />
Figure 4.1.19 gives an example of the different data structures of boundary<br />
representation and constructive solid geometry of one and the same<br />
object.<br />
4.1.5.6 Hybrid models<br />
Beside the ‘pure’ representations of object models often hybrid models are<br />
utilized. Hereby the term ‘hybrid model’ is not clearly defined and is used<br />
for various types of models, which use different representations within one<br />
system. In principle the term can be used for all non-homogeneous models.<br />
It is often used for systems that can handle wireframe and surface<br />
models at the same time. Others use the term ‘hybrid model’ for using<br />
volume and surface models under the same graphical user interface. Or<br />
hybrid models are volume models, which use partly a CSG as well as a<br />
Figure 4.1.19 BRep and CSG data structure of the same object.
332 Pierre Grussenmeyer et al.<br />
BRep data structure in order to have a dual representation of generative<br />
and accumulative elements. Hybrid models distinguish a primary structure,<br />
which is responsible for the exact model and all relevant model<br />
algorithms, and a secondary structure for specific tasks. The main task of<br />
this secondary structure is often the fast visualization of objects on the<br />
graphical screen.<br />
4.1.6 Visual reality<br />
Due to the progress in computer hard- and software there is a rapid development<br />
in the facilities of visualization in architectural photogrammetry.<br />
Simple facade plans are no longer suitable for the demands and applications<br />
of many users. 3D-real-time applications such as animations, interactive<br />
fly-overs and walk-arounds, which had needed the performance of<br />
high-end workstations a few years ago, are now also available on personal<br />
computers.<br />
Two different concepts have to be distinguished. Where ‘Virtual Reality’<br />
mainly uses vector models to describe a non-existing ( virtual) situation<br />
or fiction, ‘Visual Reality’ means a complex combination of vectors,<br />
surfaces and photo textures to visualize an existing object (‘photomodel’).<br />
A fascinating idea is to merge these models into one, showing virtual<br />
objects within a visualized reality.<br />
Figure 4.1.20 Perspective view of 3D photomodel ‘Ottoburg, Innsbruck’.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 333<br />
Figure 4.1.21 Orthoimages in scale of 3 facades of photomodel ‘Ottoburg,<br />
Innsbruck’.<br />
According to the purpose and required accuracy of the result there are<br />
numerous ways to create such textured 3D models. They range from<br />
sticking ortho-rectified photos to geometrically simplified surfaces of<br />
facades to a sophisticated reprojection of the original photos on to the<br />
complex geometry of a building using interior and exterior orientation of<br />
the camera.<br />
Regarding these models of a monument’s 3D data as a basic storage<br />
concept, a large number of resulting products can be derived from it. As<br />
examples arbitrary perspective views and orthoimages in scale should be<br />
referenced here.<br />
A very promising way to visualize 3D-data is to create so-called ‘worlds’,<br />
not only for computer games but also for ‘more serious’ applications. VRML<br />
is a new standardized format (ISO 1997) describing three-dimensional<br />
models and scenes including static and dynamic multi-media elements. This<br />
description format is independent of the kind of computer. Most Internet<br />
browsers support VRML file format. 3D object models can be viewed and<br />
inspected interactively by the user or animated in real-time even on a PC.<br />
Thus, VRML is well suited to create, for example, interactive environments,<br />
virtual museums, visualizations and simulation based on real-world data.<br />
Another way to visualize real-world objects is creating panoramic images.<br />
This approach avoids the time-consuming process needed for a 3D model.<br />
Plug-ins for Web browsers provide interactive movement. There are several<br />
methods to achieve panoramic images. One is to take single images with<br />
20 per cent to 50 per cent overlap from a fixed position while rotating<br />
the camera around a vertical axis. Warping them on to a cylindrical or<br />
spherical surface leads to a spatial imagination when navigating through<br />
the model. Another way is to move the camera around the object with a<br />
fixed target point. Complex objects can be so viewed and turned around<br />
on a personal computer by simply dragging the mouse.
334 Pierre Grussenmeyer et al.<br />
A combination of multiple panoramas or linking additional information<br />
about the shown objects can be done using clickable ‘hot spots’. Special<br />
panoramic cameras and authoring tools for image stitching are available.<br />
The new image-based rendering techniques using synthesis of images<br />
(among others to create panoramic views) also work without explicit 3D<br />
models of the object. Parallaxes between two images suffice for interpolation<br />
(sometimes even extrapolation) of others. They are restricted to<br />
existing objects and cannot be to combined with virtual worlds.<br />
4.1.7 International Committee for Architectural<br />
<strong>Photogrammetry</strong> (CIPA)<br />
To conclude this chapter in which an overview of recent developments<br />
and applications in architectural photogrammetry are given, we briefly<br />
present the International Committee for Architectural <strong>Photogrammetry</strong><br />
(CIPA), as a forum on this field.<br />
CIPA is one of the international committees of ICOMOS (International<br />
Council on Monuments and Sites) and it was established in collaboration<br />
with ISPRS (International Society of <strong>Photogrammetry</strong> and Remote Sensing).<br />
Its main purpose is the improvement of all methods for the surveying of<br />
cultural monuments and sites, specially by synergy effects gained by the<br />
combination of methods under special consideration of photogrammetry<br />
with all its aspects, as an important contribution to the recording and perceptual<br />
monitoring of cultural heritage, to the preservation and restoration<br />
of any valuable architectural or other cultural monument, object or site, as<br />
a support to architectural, archaeological and other art-historical research.<br />
ISPRS and ICOMOS created CIPA because they both believe that a<br />
monument can be restored and protected only when it has been fully<br />
measured and documented and when its development has been documented<br />
again and again, i.e. monitored, also with respect to its environment, and<br />
stored in proper heritage information and management systems.<br />
In order to accomplish this mission, CIPA [see http://www.cipa.uibk.<br />
ac.at] will:<br />
• establish links between architects, historians, archaeologists, conservationists,<br />
inventory experts and specialists in photogrammetry and<br />
remote sensing, spatial information systems, CAAD, computer graphics<br />
and other related fields;<br />
• organize and encourage the dissemination and exchange of ideas,<br />
knowledge, experience and the results of research and development<br />
(CIPA Expert Groups and CIPA Mailing List);<br />
• establish contacts with and between the relevant institutions and<br />
companies which specialize in the execution of photogrammetric<br />
surveys or in the manufacture of appropriate systems and instruments<br />
(Board of Sustaining Members);
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
• initiate and organize conferences, symposia, specialized colloquia,<br />
workshops, tutorials, practical sessions and specialized courses (CIPA<br />
Events);<br />
• initiate and coordinate applied research and development activities<br />
(CIPA Working Groups);<br />
• undertake the role of scientific and technical expert for specific projects<br />
(CIPA Expert Advisory Board);<br />
• organize a network of National and Committee Delegates;<br />
• submit an annual report on its activities to the ICOMOS Bureau<br />
(Secretary General) and the ISPRS Council (Secretary General) and<br />
publish it on the Internet (Annual Reports);<br />
• publish also its Structure, its Statutes and Guidelines on the Internet.<br />
CIPA has a well-established structure of Working Groups (WG) and<br />
Task Groups (TG):<br />
• WG 1 – Recording, Documentation and Information Management<br />
Principles and Practices;<br />
• WG 2 – Cultural Heritage Information Systems;<br />
• WG 3 – Simple Methods for Architectural <strong>Photogrammetry</strong>;<br />
• WG 4 – <strong>Digital</strong> Image Processing;<br />
• WG 5 – Archaeology and <strong>Photogrammetry</strong>;<br />
• WG 6 – Surveying Methods for Heritage Recorders;<br />
• WG 7 – Photography;<br />
• WG 8 – Cultural Landscapes;<br />
• TG 1 – Non-professional Heritage Recorders;<br />
• TG 2 – Single Images in Conservation.<br />
Bibliography<br />
References from books<br />
Architectural photogrammetry 335<br />
Atkinson, K.B., 1996. Close Range <strong>Photogrammetry</strong> and Machine Vision. Whittles<br />
Publishing, London.<br />
Batic J., et al., 1996. <strong>Photogrammetry</strong> as a Method of Documenting the Cultural<br />
Heritage (in English and Slovenian). Ministry of Culture, Ljubljana, Slovenia.<br />
Dallas, R.W.A., 1996. Architectural and archaeological photogrammetry. In K.B.<br />
Atkinson (ed.) Close Range <strong>Photogrammetry</strong> and Machine Vision. Whittles Publishing,<br />
Caithness, UK, pp. 283–302.<br />
Fondelli, M., 1992. Trattato di fotogrammetria urbana e architettonica (in Italian).<br />
Gius Laterza and Figli Spa (eds), Roma-Bari, Italy.<br />
Grätz, J.F., 1989. Handbuch der 3D-CAD-Technik. Modellierung mit 3D-<br />
Volumensystemen (in German). Siemens-Aktiengesellschaft, Berlin.<br />
Kraus, K. with contributions by Waldheusl P., 1993. <strong>Photogrammetry</strong>, vol. 1,<br />
Fundamentals and Standard Processes, 4th edition, Dümmler, Bonn, ISBN<br />
3–427–78684–6. Translated into French by P. Grussenmeyer and O. Reis 1998
336 Pierre Grussenmeyer et al.<br />
Manuel de photogrammétrie – principes et procédés fondamentaux, Editions<br />
HERMES, Paris.<br />
Kraus, K. with contributions by J. Jansa and H. Kager, 1997. <strong>Photogrammetry</strong> –<br />
Advanced Methods and Applications. vol. 2, 4th edition, Dümmler, Bonn.<br />
Luhmann, Th., 2000. Nahbereichsphotogrammetrie – Grundlagen, Methoden und<br />
Anwendungen (in German). Wichmann-Verlag, Heidelberg.<br />
Mäntylä, M., 1988. Introduction to Solid Modelling. Computer Science Press,<br />
Maryland, US.<br />
Patlas, P., Karras, G.E., 1995. Contemporary Photogrammetric Applications in<br />
Architecture and Archaeology (in Greek). University of Thessaloniki.<br />
Saint-Aubin, J.-P., 1992. Le relevé et la représentation de l’architecture (in French).<br />
Inventaire Général des Monuments et des Richesses Artistiques de la France,<br />
Paris, 232pp. (in French).<br />
Weimann, G., 1988. Architektur-Photogrammetrie (in German). Wichmann Verlag,<br />
Karlsruhe, Germany.<br />
References from journals and other literature<br />
Agnard, J.-P., Gagnon, P.-A., Nolette, C., 1988. Microcomputers and photogrammetry.<br />
A new tool. The Videoplotter. PEandRS, 54 (8), pp. 1165–1167.<br />
Agnard, J.P., Gravel, C., Gagnon, P.-A., 1998. Realization of a digital phototheodolite.<br />
International Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol.<br />
XXXII, Part 5, Hakodate, pp. 498–501.<br />
Almagro, A., 1999. <strong>Photogrammetry</strong> for everybody. International Archives of<br />
<strong>Photogrammetry</strong> and Remote Sensing, vol. XXXII, CIPA Symposium, Olinda,<br />
Brazil.<br />
Baltsavias, E., Bill, R., 1994. Scanners – a survey of current technology and future<br />
needs. International Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol. 30,<br />
Part 1, pp. 130–143.<br />
Baltsavias, E., Waegli, B., 1996. Quality analysis and calibration of DTP scanners.<br />
International Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol. 31, Part<br />
B1, pp. 13–19.<br />
Bryan, P.G., Corner, I., Stevens, D., 1999. <strong>Digital</strong> rectification techniques for architectural<br />
and archeological presentation. Photogrammetric Record, 16(93) (April),<br />
pp. 399–415.<br />
Chengshuang, L., Rodehorst, V. Wiedemann, A., 1997. <strong>Digital</strong> image processing<br />
for automation in architectural photogrammetry. In O. Altan and L. Gründig<br />
(eds) Second Turkish-German Joint Geodetic Days. Berlin, Germany, May<br />
28–30, 1997. Istanbul Technical University, pp. 541–548.<br />
CIPA, 1999. Questionnaire on the processing of the data set ‘Zurich city hall’.<br />
Edited by CIPA Working Group 3 and 4 (A. Streilein, P. Grussenmeyer and K.<br />
Hanke), 1999. 8 pp. Available at http://www.cipa.uikk.ac.at<br />
Dallas, R.W.A., Kerr, J.B., Lunnon, S., Bryan, P.G., 1995. Windsor Castle: photogrammetric<br />
and archaelogical recording after the fire. Photogrammetric Record,<br />
15(86), pp. 225–240.<br />
Drap, P., Grussenmeyer, P., 2000. A digital photogrammetric workstation on the<br />
WEB. ISPRS Journal of <strong>Photogrammetry</strong> and Remote Sensing, 55(1), pp. 48–58.<br />
Egels, Y., 1998. Monuments historiques et levers photogrammétriques. Revue Géomètre,<br />
(3), pp. 41–43, in French.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 337<br />
El-Hakim, S., 2000. A practical approach to creating precise and detailed 3D<br />
models from single and multiple views. International Archives of <strong>Photogrammetry</strong><br />
and Remote Sensing, vol. 33, Part 5, pp. 203–210.<br />
Fellbaum, M., 1992. Low-cost systems in architectural photogrammetry. International<br />
Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol. XXIX, Part B5,<br />
Washington DC, pp. 771–777.<br />
Glykos, T., Karras, G.E., Voulgaridis, G., 1999. Close-range photogrammetry<br />
within a commercial CAD package. International Archives of <strong>Photogrammetry</strong><br />
and Remote Sensing, vol. XXXII, Part 5W11, Thessaloniki, pp. 103–106.<br />
Gruen, A., 1976. Photogrammetrische Rekonstruktion aus Amateuraufnahmen.<br />
Architektur-Photogrammetrie II, Arbeitsheft 17, Landeskonservator Rheinland.<br />
Rheinland-Verlag, Köln, pp. 85–92.<br />
Gruen, A., 2000. Semi-automated approaches to site recording and modelling.<br />
International Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol. 33 Part 5,<br />
pp. 309–318.<br />
Grussenmeyer, P., Abdallah, T., 1997. The cooperation in architectural photogrammetry<br />
between ENSAIS (France) and ECAE (Egypt): practical experiences on<br />
historic monuments in CAIRO. International Archives of <strong>Photogrammetry</strong> and<br />
Remote Sensing, vol. XXXII, Part 5C1B, Göteborg, pp. 215–221.<br />
Grussenmeyer, P., Guntz, C., 1997. Photogrammétrie architecturale et réalité<br />
virtuelle: modélisation de l’aqueduc El-Ghuri (Le Caire, Egypte). Revue de<br />
l’Association Française de Topographie, XYZ 4 e trim. 97 no. 73, pp. 75–81, in<br />
French.<br />
Grussenmeyer, P., Koehl, M., 1998. Architectural photogrammetry with the<br />
TIPHON software, towards digital documentation in the field. International<br />
Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol. XXXII, Part 5, Hakodate,<br />
pp. 549–556.<br />
Hanke, K., 1994. The Photo-CD – A Source and <strong>Digital</strong> Memory for Photogrammetric<br />
Images. International Archives of <strong>Photogrammetry</strong> and Remote<br />
Sensing. vol. XXX Part 5, Melbourne, 1994, pp. 144–149.<br />
Hanke, K., 1998. <strong>Digital</strong> close-range photogrammetry using CAD and raytracing<br />
techniques. International Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol.<br />
XXXII, Part 5, Hakodate, pp. 221–225.<br />
Hanke, K., Ebrahim, M.A-B., 1997. A low cost 3D-measurement tool for architectural<br />
and archaeological applications. International Archives of <strong>Photogrammetry</strong><br />
and Remote Sensing, vol. XXXI, Part 5C1B, CIPA Symposium,<br />
Göteborg 1997, pp. 113–120.<br />
Hanke, K., Ebrahim, M.A-B., 1999. The ‘digital projector’: Raytracing as a tool<br />
for digital close-range photogrammetry. ISPRS Journal of <strong>Photogrammetry</strong> and<br />
Remote Sensing, 54(1), Elsevier Science B.V., Amsterdam, pp. 35–40.<br />
Hemmleb, M., Wiedemann A., 1997. <strong>Digital</strong> Rectification and Generation of Orthoimages<br />
in Architectural <strong>Photogrammetry</strong>. International Archives of <strong>Photogrammetry</strong><br />
and Remote Sensing vol. XXXI Part 5C1B, CIPA Symposium, Göteborg,<br />
pp. 261–267.<br />
Ioannidis, C., Potsiou, C., Badekas, J., 1996. 3D detailed reconstruction of a<br />
demolished building by using old photographs. International Archives of<br />
<strong>Photogrammetry</strong> and Remote Sensing, vol. XXXI, Part B5, Vienna,<br />
pp. 16–21.
338 Pierre Grussenmeyer et al.<br />
Jachimski, J., 1995. Video stereo digitizer. A small digital stereophotogrammetric<br />
working station for the needs of SIT and other application. Polish Academy of<br />
Sciences. The Krakow Section. Proceedings of the Geodesy and Environmental<br />
Engineering Commission, Geodesy 38, pp. 71–91.<br />
Karras, G., Patlas, P., Petsa, E., Ketipis, K., 1997. Raster projection and development<br />
of curved surfaces. International Archives of <strong>Photogrammetry</strong> and Remote<br />
Sensing, vol. XXXII, Part 5C1B, Göteborg, pp. 179–185.<br />
Kasser, M., 1998. Développement d’un photothéodolite pour les levés archéologiques.<br />
Revue Géomètre, (3), pp. 44–45, in French.<br />
Lenz, R., Lenz, U., 1993. New development in high resolution image acquisition<br />
with CCD-array sensor. In A. Gruen and H. Kahmen (eds) Optical 3-D<br />
Measurement Techniques II. Wichmann-Verlag, Karlsruhe, 1993. S. 53–62.<br />
Patias, P., Peipe, J., 2000. <strong>Photogrammetry</strong> and CAD/CAM in culture and industry,<br />
an ever-changing paradigm. International Archives of <strong>Photogrammetry</strong> and<br />
Remote Sensing, vol. 33 Part 5, pp. 599–603.<br />
Patias, P., Streilein, A., 1996. Contribution of videogrammetry to the architectural<br />
restitution – Results of the CIPA ‘O. Wagner Pavillon’ test. International Archives<br />
of <strong>Photogrammetry</strong> and Remote Sensing, vol. XXI, Part B5, Vienna, pp. 457–462.<br />
PCWebopaedia, 2000. <strong>Digital</strong> camera. URL: www.pcwebopaedia.com/digital_<br />
camera.htm. (accessed November 2000).<br />
Plugers, P., 1999. Product survey on digital photogrammetric workstations. GIM<br />
International, May, pp. 61–65.<br />
Plug-In Systems, 2000. The digital camera guide. URL: www.pluginsystems.com.<br />
(accessed November 2000).<br />
Pomaska, G., 1998. Automated processing of digital image data in architectural<br />
surveying. International Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol.<br />
XXII, Part 5, Hakodate, pp. 637–642.<br />
Reulke, R., Scheele, M., 1997. CCD-line digital imager for photogrammetry in<br />
architecture. International Archives of <strong>Photogrammetry</strong> and Remote Sensing,<br />
vol. XXXII, Part 5C1B, Göteburg, pp. 195–201.<br />
Schneider, C-T., 1996. DPA-WIN – A PC-based digital photogrammetric station<br />
for fast and flexible on-site measurement. International Archives of <strong>Photogrammetry</strong><br />
and Remote Sensing, vol. XXI, Part B5, Vienna, pp. 530–533.<br />
Streilein, A., 1995. Videogrammetry and CAAD for architectural restitution of the<br />
Otto-Wagner-Pavilion in Vienna. In A. Gruen and H. Kahmen (eds) Optical 3-<br />
D Measurement Techniques III, Wichmann-Verlag, Heidelberg, pp. 305–314.<br />
Streilein, A., 1996. Utilization of CAD models for the object oriented measurement<br />
of industrial and architectural objects. International Archives of <strong>Photogrammetry</strong><br />
and Remote Sensing, vol. XXI, Part B5, Vienna, pp. 548–553.<br />
Streilein, A., Beyer, H., Kersten, T., 1992. <strong>Digital</strong> photogrammetric techniques for<br />
architectural design. International Archives of <strong>Photogrammetry</strong> and Remote<br />
Sensing, vol. XXIX, Part B5, Washinton DC, pp. 825–831.<br />
Streilein, A., Gaschen, S., 1994. Comparison of a S-VHS camcorder and a highresolution<br />
CCD-camera for use in architectural photogrammetry. International<br />
Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol. XXX, Part 5, Melbourne,<br />
pp. 382–389.<br />
Streilein, A, Hanke, K., Grussenmeyer, P., 2000. First experiences with the ‘Zurich<br />
City Hall’ data set for architectural photogrammetry. International Archives of<br />
<strong>Photogrammetry</strong> and Remote Sensing, vol. 33 Part 5, pp. 772–779.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Architectural photogrammetry 339<br />
Streilein, A., Niederöst, M., 1998. Reconstruction of the Disentis monastery from<br />
high-resolution still video imagery with object-oriented measurement routines.<br />
International Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol. XXII, Part<br />
5, Hakodate, pp. 271–277.<br />
Thomas, P.R., Miller, J.P., Newton, I., 1995. An investigation into the use of<br />
Kodak Photo-CD for digital photogrammetry. Photogrammetric Record 15(86),<br />
pp. 301–314.<br />
Van den Heuvel, F.A., 1998, 3D reconstruction from a single image using geometric<br />
constraints. ISPRS Journal of <strong>Photogrammetry</strong> and Remote Sensing, 53(6),<br />
pp. 354–368.<br />
Van den Heuvel, F.A., 1999. Estimation of interior orientation parameters from<br />
constraints on line measurements in a single image. International Archives of<br />
<strong>Photogrammetry</strong> and Remote Sensing, vol. 32, Part 5W11, Thessaloniki, pp.<br />
81–88.<br />
Varhosaz, M., Dowman, I., Chapman, D., 2000. Towards automatic reconstruction<br />
of visually realistic models of buildings. International Archives of <strong>Photogrammetry</strong><br />
and Remote Sensing, vol. 33 Part 5, pp. 180–186.<br />
Waldheusl, P., 1992. Defining the future of architectural photogrammetry.<br />
International Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol. XXX, Part<br />
B5, Washington DC, pp. 767–770.<br />
Waldheusl, P., 1999. Tasks for ISPRS working groups to serve ICOMOS.<br />
International Archives of <strong>Photogrammetry</strong> and Remote Sensing, vol. XXXII,<br />
Part 5W11, Thessaloniki, pp. 1–7.<br />
Waldheusl, P., Brunner, M., 1988. Architectural photogrammetry world-wide and<br />
by anybody with non-metric cameras? Proceedings des XI. Internationalen CIPA<br />
Symposiums, October, Sofia, pp. 35–49.<br />
Waldheusl, P.,Ogleby, C., 1994. 3 × 3-Rules for simple photogrammetric documentation<br />
of architecture. International Archives of <strong>Photogrammetry</strong> and Remote<br />
Sensing, vol. XXX, Part 5, Melbourne, pp. 426–429.<br />
Wester-Ebbinghaus, W., 1989. Das Réseau im photogrammetrischen Bildraum.<br />
Bildmessung und Luftbildwesen, 3(89), pp. 64–71.<br />
Wiedemann, A., 1996. <strong>Digital</strong> orthoimages in architectural photogrammetry using<br />
digital surface models. International Archives of <strong>Photogrammetry</strong> and Remote<br />
Sensing, vol. XXXI, Part B5, Vienna, pp. 605–609.<br />
Wiedemann, A., Hemmleb, M., Albertz, J., 2000. Recontruction of historical buildings<br />
based on images from the Meydenbauer archives. International Archives of<br />
<strong>Photogrammetry</strong> and Remote Sensing, vol. 33 Part 5, pp. 887–893.<br />
Wiedmann, A., Rodehorst, V., 1997. Towards automation in architectural photogrammetry<br />
using digital image processing. International Archives of <strong>Photogrammetry</strong><br />
and Remote Sensing, vol. XXXI, Part 5C1B, CIPA Symposium, Göteborg,<br />
pp. 209–214.<br />
Zischinsky, T., Dorffner, L., Rottensteiner, F., 2000. Application of a new model<br />
helicopter system in architectural photogrammetry. International Archives of<br />
<strong>Photogrammetry</strong> and Remote Sensing, vol. 33 Part 5, pp. 959–968.
340 Michel Kasser<br />
4.2 PHOTOGRAMMETRIC METROLOGY<br />
Michel Kasser<br />
4.2.1 Introduction<br />
The term ‘photogrammetric metrology’ covers the whole range of metrology<br />
activities that exploit photogrammetric processes, that is to say,<br />
geometric processes based on image acquisition, and the image processing<br />
which historically hardly ever took place in real time. The appearance of<br />
digital imagery has been the origin of a significant change in this technical<br />
domain, especially because it allowed real-time measures to be reached,<br />
which opened many new markets. And then the use of CCD cameras also<br />
permitted images of very large dynamics to be obtained, allowing the difficult<br />
cases formed by uniform surfaces, very current in industrial metrology,<br />
to be processed in a much better way. The very large dynamics makes it<br />
possible to detect very weak differences of radiometry having a meaningful<br />
character, which means they can be processed by automatic correlation.<br />
Let us recall that photogrammetry is fundamentally a method founded<br />
on angle measures, angles that are recorded permanently on the film or<br />
in the digital pixel geometry. From this point of view it is easy to make<br />
the link with metrology methods based on theodolites measures, which<br />
are also used to measure angles, but with quite different devices. Therefore<br />
there is a continuity of methodologies between these two domains, and<br />
now that theodolites are digital, sometimes motorized, or capable of<br />
pointing their targets automatically, one notes without surprise that the<br />
same softwares are often used to process measures coming from these two<br />
techniques.<br />
There is also a continuity between photogrammetry and studies in<br />
robotics to permit a fast identification of the geometric environment of<br />
the robot. The robotics developed some continual measures based on video<br />
images, allowing some elements in an image to be treated, this being<br />
performed at very high speed in order to allow the calculator to choose<br />
movements that the different motors have to perform. It is still photogrammetry,<br />
but a very specialized one, and that one sometimes calls<br />
videogrammetry.<br />
Photogrammetric metrology finally represents all the methods using the<br />
equation of collinearity (see §1.1) on images, but in conditions that differ<br />
from the simple field of cartography. It represents a reuse of the body of<br />
knowledge of photogrammetry, but for applications that are each different,<br />
and that are rather specialized towards measures without contact with a<br />
large amounts of points. Objects processed are generally of restricted sizes,<br />
from about ten metres (wing of plane, cockles of boat) down to the<br />
millimetre, or even less (inspection of tapped holes, of soldering, of surfaces<br />
of samples checked with a microscope, etc.).
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
4.2.2 Equipment<br />
Photogrammetric metrology 341<br />
The basis of modern digital photogrammetric metrology is the CCD<br />
camera. This type of sensor has been used with all available matrix sizes,<br />
from small-size matrixes for real-time applications up to 4k × 4k matrixes,<br />
and probably more sooner or later. This matrix is set up as usual in the<br />
focal plane of the optics, and as it is about doing measures these optics<br />
must have some steady geometric parameters in the time, so that they can<br />
make the object of a calibration. In this sense one doesn’t search now, as<br />
in aerial imagery (see Chapter 2), for optics without distortion, but rather<br />
for optics having an excellent resolution, and whose distortion is steady<br />
in time, and the same for the position of their optic centre and their principal<br />
planes.<br />
This general type (CCD optics) can be used in very different ways. It<br />
can first of all be used locally, like a traditional camera, with successive<br />
positions of the same device or several synchronized devices, and while<br />
identifying the points in images on which are done the external geometric<br />
measures. The setting up of images is then achieved using traditional tools<br />
of photogrammetry, with possible adaptations for geometries very different<br />
from the aerial images with an almost vertical axis (e.g. problems of<br />
refraction).<br />
It can also be used fixed very rigidly to the optic axis of a theodolite,<br />
which permits one to know the orientation in the reference frame corresponding<br />
to every image with the considerable precision of the system<br />
of angles measures of the theodolite. This whole is named phototheodolite;<br />
it was used a long time ago with film or even glass plates cameras, but<br />
has since fallen into obsolescence. If the theodolite is itself localized and<br />
oriented, which is a simple operation, one gets images that possess all<br />
parameters of localization and orientation, permitting an immediate setting<br />
up to do directly photogrammetric restitution. One can also use only the<br />
capacity of orientation of the theodolite to combine one set of images<br />
obtained in the different directions (motorized phototheodolite), and to<br />
get thus in a unique conical geometry an image having a very important<br />
number of pixels.<br />
One can also try to measure rapidly a large number of targets (for only<br />
a few, a motorized theodolite with automatic aiming would be preferable),<br />
while going very quickly, typically in some milliseconds. In this case one<br />
uses retroreflecting targets, formed from many small plastic cubes’ corners<br />
for example, and the optics used is then also equipped with an annularshaped<br />
flash, which forms a circle completely surrounding the frontal lens.<br />
Thus targets send back a considerable light quantity towards the objective,<br />
and in return with a suitable threshold of the image one will see on<br />
this one only luminous points on a black scene, every point corresponding<br />
to a target. It will remain to identify these different targets on every acquired<br />
image, which requires an approximate knowledge of the measured object.
342 Michel Kasser<br />
4.2.3 Cases of use<br />
The specificities of photogrammetric metrology must be looked for in relation<br />
to other methods of measure, in order to identify its typical uses. We<br />
may consider it especially useful in the following domains:<br />
• when the measure must be acquired in a very short time on numerous<br />
to very numerous points (from tens to hundreds of thousands of<br />
points);<br />
• when measures are to be performed without possible or desirable<br />
contact (very hot objects, contaminated objects, objects for which no<br />
pose of targets is admissible);<br />
• when measures can be exploited only a long time after the acquisition<br />
of the images: the image acquisition is an inexpensive part of the<br />
complete photogrammetric operation, and image processes can be put<br />
back to a later period, for example to measure the ageing of a given<br />
object.<br />
We will examine here some examples concerning one or several of these<br />
domains.<br />
Medical domain<br />
Two situations are typical of this domain: the follow-up geometric of a<br />
portion of the human body (lesions, tumours, dermatological problems,<br />
but also identification of people by morphometric measures . . .), and the<br />
surgical intervention from afar allowing a specialist to intervene quickly<br />
on distant sites without displacing the patient or the surgeon. The second<br />
one is again in development, but it is likely that it will be limited to<br />
transmitting some stereoscopic images from a site equipped for a telemanipulation<br />
of surgical instruments, and that photogrammetric aspects<br />
will be quite reduced (possible 3D measures from afar). The first requires<br />
more classic concepts. Being about a living surface, it will be necessary to<br />
proceed to an instantaneous measure of two images forming a stereoscopic<br />
couple. And as it is not simple to put in the field of every image some<br />
elements permitting a formation of the model in differed time, the most<br />
logical answer is to use two cameras fixed very rigidly in relation to each<br />
other. Thus, once calibration measures have been achieved, one is able to<br />
measure in 3D the part of the body appearing in the zone covered in stereoscopy.<br />
An aid by automatic correlation gives assistance in pointing for<br />
non-photogrammetres practitioners. However, as we will show in §4.2.4,<br />
the extremely low variable radiometry of human skin requires a special<br />
lighting so that the whole functions correctly.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Photogrammetric metrology 343<br />
Objects in movement<br />
This is quite a tyical case for the use of photogrammetry. We are in the<br />
presence of an object that distorts itself quickly, or of an object in movement,<br />
and the number of points to measure is superior to a few units (if<br />
the number is very low it can be performed with motorized automatic<br />
pointing theodolites). Examples of such situations are numerous: study of<br />
pieces distorting themselves under strain (tests of plane wings), dynamic<br />
phenomenon studies (survey of the wave formation to the stem of a ship,<br />
fall of blocks in an unsteady slope), etc. The images are in this case<br />
necessarily acquired with several cameras, which must be synchronized<br />
rigorously. The geometric elements allowing the setting up to be achieved<br />
are obtained according to what is feasible, every situation forming a special<br />
case: some known coordinate targets are in the field of cameras, or again<br />
localization and orientation in the space of cameras are obtained by external<br />
means (GPS, inertial systems, goniometric orientation . . .).<br />
Controls of coppercraft pieces<br />
Pieces of thick coppercraft cause subtle metrological problems: it is generally<br />
about the control of the obtained shape in compliance with the initial<br />
specifications. Some pieces of very large size can require that the reality<br />
be very close to the surfaces calculated (propellers of ships, wings of<br />
plane, parabolic reflectors for radio waves for telecommunications, radioastronomy,<br />
etc.). The most economic solution here is still photogrammetry,<br />
which permits the requested precision to be reached (often better than<br />
1 mm) in a much more cost-effective way than other methods. The other<br />
main technical solution consists in using some tridimensional measuring<br />
machines, but these machines are limited in accessible measurements (the<br />
largest ones don’t permit one to make measurements of objects larger than<br />
3 m), and are slow (several seconds between every measured point), but<br />
are on the other hand much more precise (they measure with a precision<br />
of the order of the m). Besides, they are extremely costly, and pieces<br />
must be brought to the machine, and not the reverse. <strong>Photogrammetry</strong>,<br />
although not so precise, does not present these limitations. Besides, the<br />
acquisition of the whole image can take place in a very short time, as we<br />
already saw. Without reaching the complete simultaneity, which requires<br />
arranging numerous cameras, it is possible to acquire the whole image if<br />
need be in less than one hour. This type of possibility is profitable in the<br />
case of large-sized objects for which the temperature creates distortions,<br />
that cannot be controlled (objects installed outside, as for parabolic<br />
antennas, ship sections intended to be fixed together, etc.).
344 Michel Kasser<br />
Surveys of painted underground caves<br />
We are here in the case of artistic productions, onto which it is sometimes<br />
completely impossible to put targets considering the fragility of the object.<br />
<strong>Photogrammetry</strong> permits measures without contact, without any risk of<br />
destruction for the studied object. In counterpart, one sometimes expects<br />
in such cases the possibility of restoring the acquired images under a plane<br />
shape, or developed on a mean plane, which is evidently not possible<br />
without the calculation of a complete 3D model of this object. So a fresco<br />
achieved on a surface that includes numerous mouldings, or tracings of<br />
animals on prehistoric underground caves, are cases in which elements of<br />
reliefs of the cave walls may be an integral part of the tracing and are<br />
therefore indispensable for the archaeologists’ interpretation. In such situations,<br />
it is necessary to try to identify some precise and punctual details<br />
in sufficient quantities that will be determined by classic topometric<br />
methods. If this is not possible, one can create some relatively precise<br />
temporary details while using low-power visible laser beams (red laser<br />
diode pointers, or impacts of reflectorless laser EDMs, of Disto type (Leica)<br />
for example). There is also the possibility of working with phototheodolites,<br />
which can be used in order to deliver oriented and localized images<br />
with a high precision. It allows images to be restored without using elements<br />
of location in the images themselves. The inconvenience of this solution<br />
resides in the impossibility of the operator guaranteeing himself against<br />
a lawlessness of the material under utilization. For example, if the optic<br />
axes of the theodolite and the camera have relative directions that<br />
change without the knowledge of the operator (shocks, vibrations), it will<br />
a posteriori induce mistakes that one will often not be able to correct.<br />
It is then veritably a blind work.<br />
Surveys of high-temperature objects<br />
The possibility of measuring with absolutely no contact with the measured<br />
object is therefore a classic field of the application of photogrammetry. So<br />
to control streams of incandescent metal, there is virtually no equivalent<br />
solution. Cameras are then adapted to such environments, but processes<br />
done on images are always of photogrammetric type. One benefits then<br />
from the capacity to measure without contact, as well as the one to acquire<br />
images instantaneously. Two stationary, rigid cameras can be installed<br />
permanently, and their relative positions will be the topics of a complete<br />
calibration by taking images of a tridimensional object including targets,<br />
whose relative positions will have been measured with a very high precision.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Photogrammetric metrology 345<br />
4.2.4 Typical problems of photogrammetric metrology<br />
Photogrammetric metrology may be adapted, as a question of principle,<br />
to situations that change often, and at each time it must be adjusted to<br />
the different cases, as we have seen previously. Nevertheless, certain technical<br />
problems are recurrent, because they are bound to the same bases<br />
of the photogrammetric process. We will present here certain among them.<br />
Case of uniform surfaces<br />
This case is often encountered: a uniform surface will be defined here as<br />
a surface whose apparent radiometry varies very little on an important<br />
area. This very weak variation itself is a function of two parameters: the<br />
dynamics of the sensor, and the real variation of the radiometry. The<br />
reason why this dynamics is very important is it becomes possible to<br />
measure some extremely reduced radiometry differences, and thus the<br />
uniform zones become rarer and in any case of more reduced size than<br />
with a more modest dynamics sensor. Surfaces of nearly uniform radiometry<br />
are very frequent, and are met with in most of the previously evoked<br />
cases. So in medical imagery, they are surfaces of skin, or in coppercraft,<br />
surfaces of raw metal. On such surfaces it is then impossible to point with<br />
the eye, and tools of automatic correlation may not be able to provide<br />
some satisfactory indications. One solution may prove to be satisfactory<br />
in such cases: it consists in using a special lighting if the situation permits<br />
it. This lighting is chosen in order to project a whole set of random spots<br />
on the object. For example, for a metric dimension object one will be able<br />
to use a lighting provided by a retroprojector, on the tray of which<br />
will have been arranged a film producing such very tight random spots.<br />
Of course it will be necessary that the intended motive does not include<br />
any periodicity, otherwise it will become very difficult to avoid false<br />
correlations.<br />
Case of reflecting surfaces<br />
In the same way it will be impossible to process correctly by photogrammetry<br />
reflecting surfaces without making a particular preparation of the<br />
object to process. The solution proposed most frequently consists in such<br />
cases (polite metal, glazing, etc.) of putting down a thin talc deposit, and<br />
there again illuminating the surface while projecting some random spots.<br />
The deposit of talc is easy to create, and doesn’t present any difficulties<br />
being removed. But this method will not be capable of being put into operation<br />
in some situations, and it is necessary to know how to anticipate<br />
results then that will be locally mediocre (for example, liquid surface cases,<br />
or again humid surfaces in decorated underground caves).
346 Michel Kasser<br />
Numbering of targets<br />
In most areas of photogrammetry, and it is of course the case in those of<br />
metrology for which one searches for the best possible precision, one uses<br />
targets of measurements adapted to the acquired images. These targets<br />
serve to do the aerotriangulation on the one hand (even though the term<br />
is not suited to cases of lack of aerial images, it remains used in spite of<br />
everything) and the setting up of couples. On the other hand, targets may<br />
be necessary as being points to be determined with the best precision: the<br />
pointing on targets are indeed more precise than pointing on natural<br />
surfaces. Some studies thus require a very large number of targets, and<br />
the problem is that it is necessary to identify these targets, in order to be<br />
able to find without error from one image to the other what are the homologous<br />
points. When the object to survey is very large (architectural<br />
photogrammetry, for example), the number of each target can be marked<br />
by hand with chalk. But on objects of some metres, and especially if targets<br />
are very numerous, this principle cannot be kept. One can then use labels<br />
including bar codes, capable of being read automatically at the time of<br />
the automated process of images. Nevertheless, this solution brings strong<br />
strains on the orientation of these labels, as they can become illegible when<br />
they are seen too much in perspective. When one uses classic reflecting<br />
targets and an annular flash mentioned previously, it becomes possible to<br />
localize targets automatically in the image by the use of a simple threshold.<br />
In such a case, if one provides the software with the approximate geometry<br />
of the object on which are the targets, it can itself then proceed to a<br />
numbering of targets, which allows a considerable time to be saved at the<br />
time of the exploitation of images.<br />
4.2.5 Findings, perspectives<br />
Photogrammetric metrology is a part of metrology that has still golden<br />
days ahead. It permits some types of works to be processed in a much<br />
more cost-effective way than any other methodologies. Besides, the appearance<br />
of direct CCD digital images in the 1990s, opened the doors of the<br />
real-time process. It must not be considered as a competitor with other<br />
techniques (utilization of reflectorless EDMs, of laser scanners systems, of<br />
tridimensional measuring machines). It is, for example, faster and more<br />
precise that the first, slower, less exhaustive and much more precise that<br />
the second, and for the third it is without contact, far less precise and<br />
considerably faster. It can be adjusted to objects of all sizes, and one of<br />
its essential steps (the image acquisitions) permits a storage that will allow<br />
all the processes to be repeated if need be. It allows one not to process<br />
images, therefore, and to keep them a very long time for possible future<br />
processes. Finally, let us wager that this discipline will certainly progress<br />
again, since it entered the digital era quite late. For example, one can
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Photogrammetric metrology 347<br />
expect to see phototheodolites used more and more in metrology, to provide<br />
measures of distortions of distant objects from afar.<br />
The inter-techniques collaboration between photogrammetry, image<br />
processing and robotics will probably provide many new tools in the next<br />
few years.
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
Index<br />
adaptive filtering 93<br />
adaptive sampling meshes 162<br />
additive synthesis 27<br />
aerotriangulation 6, 8, 12, 15, 62, 63,<br />
78, 116, 118, 119, 120, 123, 124,<br />
125, 126, 131, 132, 134, 135, 139,<br />
140, 149, 156, 157, 287, 299,<br />
346<br />
airborne laser ranging systems (ALRS)<br />
53, 54, 55, 56, 57, 58<br />
ambiguities: matching 170, 207;<br />
resolution of on flight 117<br />
angular fields of view 34<br />
atmospheric diffusion 16, 21, 22,<br />
35<br />
atmospheric refraction 1, 8, 12,<br />
49<br />
attenuation 17, 19, 20, 21, 22<br />
basins 255, 268, 269<br />
bundle restitution 304, 305<br />
catchment areas 258, 264, 268, 270,<br />
276, 277, 278, 281, 282<br />
CCD 1, 14, 23, 24, 34, 36, 37, 38, 39,<br />
41, 42, 43, 60, 61, 63, 66, 76, 153,<br />
168, 288, 300, 307, 311, 312, 313,<br />
314, 340, 341, 346<br />
characteristic features 251<br />
characteristic lines 164, 253, 255, 256,<br />
257, 259, 271, 275, 277, 278, 294;<br />
networks 280<br />
CIPA (International Committee for<br />
Architectural <strong>Photogrammetry</strong>)<br />
300, 315, 334, 335, 336, 337, 338,<br />
339<br />
close range photogrammetry 337<br />
collinearity 3, 13, 15, 124, 148, 155,<br />
340; equation 3, 13, 124, 155<br />
colorimetric spaces 26, 27, 32<br />
compensation of forward motion 24<br />
compression: algorithm 101, 105; by<br />
fractals 114; of images 100; rates of<br />
101, 104–6, 109, 114; reversible<br />
101, 102, 103, 106; with losses 101,<br />
104, 108<br />
constructive solid geometry 330,<br />
331<br />
contour lines 159, 160, 164, 229, 230,<br />
232, 233, 234, 239, 250, 256, 257,<br />
266, 272, 277, 292<br />
convolution operators 89<br />
correlation templates 184<br />
crests 234, 255, 256, 257, 258, 271,<br />
278, 279, 280<br />
cylindro-conical geometry 38<br />
decorrelators 105, 108<br />
Delaunay triangulation 162, 225, 229,<br />
233, 235, 251<br />
diffusion by sprays 17, 18<br />
digital cameras 24, 41, 45, 56, 64, 67,<br />
119, 124, 153, 158, 288, 294, 299,<br />
311, 312, 314, 316<br />
digital elevation models (DEM) 78,<br />
159, 160, 168, 169, 181, 195, 197,<br />
202, 208, 220, 224, 225, 226, 281,<br />
282<br />
digital photogrammetric workstations<br />
(DPW) 59, 78, 145, 156, 176, 224,<br />
300, 314, 323–4, 338<br />
digital surface models (DSM) 43,<br />
47, 49, 50, 53, 78, 159, 160,<br />
162, 164–7, 194–6, 202, 210–25,<br />
250, 288, 289, 301, 322, 323,<br />
339<br />
digital terrain models (DTM) 38, 46,<br />
52, 55–8, 62, 78, 118, 131, 152,<br />
156–60, 166, 168, 170, 177, 207,<br />
212, 220–3, 229, 251, 255, 258,<br />
261, 263–71, 277–87, 289, 294,<br />
299, 314, 317
350 Index<br />
digitization 39, 43, 58, 59, 60, 63,<br />
64–5, 68, 76, 100, 152, 158, 164,<br />
209, 253, 308, 310, 311; of aerial<br />
pictures 58<br />
discontinuities 29, 37, 61, 81, 162,<br />
187, 190, 198, 200, 202, 211, 214,<br />
216, 225, 254, 271, 280<br />
disparity 125, 131, 137, 141, 177,<br />
189, 192, 196, 200, 204, 205<br />
distortion 1, 6, 12, 13, 14, 15, 23, 24,<br />
37, 45, 63, 101, 104, 105, 108, 112,<br />
114, 173, 176, 287, 302, 303, 319,<br />
341<br />
drainage networks 257, 258, 277,<br />
281<br />
dynamic programming 200, 202, 203,<br />
205, 207<br />
dynamics 43, 44, 46, 56, 57, 65,<br />
76, 79, 81, 82, 85, 153, 340,<br />
345<br />
Earth curvature 7<br />
elastic grid surfaces 231, 243, 246<br />
entropy 103, 104, 105<br />
epipolar lines 151, 174, 175, 180, 200,<br />
202, 203, 205<br />
epipolar resampling 148, 150, 152,<br />
175, 176<br />
equalization of histograms 82<br />
false matches 188, 190, 216<br />
field curvature 23<br />
filtering 88, 89, 91, 93, 95, 99, 100,<br />
111, 112, 131, 132, 141, 143, 156,<br />
190, 217, 218, 219, 221, 224, 271,<br />
279<br />
fog 16, 17, 22, 23, 41, 44, 46<br />
forward motion compensation (FMC)<br />
36, 43, 60<br />
GALILEO 115, 124<br />
Gaussian filters 91<br />
geographic information systems (GIS)<br />
59, 155, 156, 220, 278, 285<br />
geoids 7<br />
geometric dilution of precision (GDOP)<br />
122<br />
glass plates 12, 14, 15, 62, 341<br />
global positioning system (GPS) 8, 36,<br />
37, 38, 42, 53, 54, 55, 56, 78, 115,<br />
116, 117, 118, 120, 121, 122, 123,<br />
124, 156, 285, 287, 343<br />
GLONASS (see GALILEO)<br />
GNSS 115, 124<br />
hidden parts 15, 45, 46, 57, 190, 208,<br />
222<br />
histograms of images 79; manipulations<br />
of 81<br />
homologous points 125, 128, 129,<br />
131, 132, 134, 137, 139, 140, 141,<br />
142, 143, 152, 170, 171, 173, 175,<br />
178, 179, 185, 187, 198, 202, 205,<br />
318, 322, 346<br />
homomorphic filtering 88, 95<br />
hot spots 34, 172, 289, 291, 334<br />
HSV (hue, saturation, intensity value)<br />
space 27, 28, 29, 30<br />
Huffman code 106, 107<br />
ICOMOS 334, 335, 339<br />
image pyramids 150, 320<br />
index map 78, 124, 140, 142, 143<br />
inertial measurement units (IMU)<br />
118<br />
interferogram 49, 50<br />
interferometry 47, 48, 49, 50, 52, 53,<br />
120, 164, 165<br />
join-up lines 289<br />
JPEG 102, 106, 108, 109, 112, 114,<br />
153<br />
Karhunen-Loève transform (KLT)<br />
113<br />
kinematics 118<br />
lasers 1, 53, 54, 58, 118, 120, 164,<br />
224, 225, 228, 289, 344, 346<br />
linear filtering 89, 91<br />
median filters 93, 97<br />
meshes 160, 166, 255, 260, 261, 262,<br />
263, 264, 267, 272, 281<br />
Mie diffusions 16, 17, 18<br />
multi-image texture and radiometric<br />
similarity (MITRAS) 214<br />
multi-image texture similarity (MITS)<br />
213, 214<br />
multi-image similarity function 212<br />
multiplicity 96, 131<br />
navigation 36, 54, 56, 115, 116, 118<br />
noise 1, 22, 43, 50, 61, 62, 64, 65, 66,<br />
68, 70, 71, 73, 76, 87, 88, 89, 91,<br />
93, 94, 96, 97, 98, 99, 100, 103,<br />
112, 116, 123, 136, 157, 167, 170,<br />
183, 189, 208, 209, 211, 221, 225,<br />
266, 312
1111<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
1011<br />
1<br />
2<br />
3111<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
20111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
30111<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
9<br />
40111<br />
1<br />
2<br />
3<br />
4<br />
45111<br />
official national reference frame 285<br />
orthoimages 212, 217, 289, 303<br />
orthophotography 41, 46, 56, 59, 62,<br />
118, 158–9, 221, 222, 223, 282,<br />
283, 284, 285, 286, 288, 290, 291,<br />
292, 299<br />
ortho-templates 184, 192, 197<br />
OTF (on the fly) 117, 122<br />
passages 29, 33, 98, 140, 255, 256,<br />
259, 269, 271<br />
perspective 2, 6, 13, 14, 42, 43, 115,<br />
116, 117, 120, 121, 131, 136, 137,<br />
148, 157, 184, 203, 282, 316, 317,<br />
333, 346<br />
perspective centres 115, 116, 121<br />
Photo-CD system 311<br />
photogrammetric filtering 131<br />
planarity of emulsions 14<br />
points of interest 125, 127<br />
pressurized planes 12<br />
radar 1, 47, 48, 49, 50, 51, 52, 53,<br />
120, 148, 165<br />
radarclinometry 47, 48, 50, 51<br />
radargrammetry 47, 48, 49, 50<br />
radiance 17, 21, 22, 25, 34<br />
radiometric balancing 288, 290<br />
radiometric quality 62, 63, 64, 66, 73,<br />
76, 77, 289, 310<br />
raised structures 159, 202, 220, 221,<br />
222, 224, 294<br />
raster grids (RG) 160, 161, 212<br />
Rayleigh diffusion 17, 18, 20, 22<br />
reference atmosphere 9<br />
refraction angles 10<br />
rotation matrix 2, 4, 5<br />
scanners 1, 38, 58, 59, 60, 61, 63,<br />
288, 310, 311, 314, 336, 346<br />
segmentation 34, 73, 221, 222, 223,<br />
224<br />
semi-metric cameras 308, 309, 310<br />
shape from shading 48, 50, 52, 164,<br />
165<br />
shutters 24, 36, 121<br />
similarity measures 125, 127, 169,<br />
207<br />
spectral bands 25, 33, 34, 112<br />
Index 351<br />
spectral decorrelation 112<br />
stereo matching from image space<br />
(SMI) 173, 183, 184, 186, 189–90,<br />
192, 195, 198, 202, 211, 220, 251,<br />
319<br />
stereo matching from object space<br />
(SMO) 183, 189, 190, 192, 195,<br />
197, 200<br />
stereopreparation 118, 123, 287<br />
subpixel displacement 149<br />
subtractive synthesis 27<br />
summits 2, 266, 281<br />
surface models 159, 303, 305, 328,<br />
329, 330, 331<br />
systematism 15, 66, 123, 124, 127,<br />
257<br />
terrestrial photogrammetry 10, 11, 14,<br />
137, 162, 300<br />
thalwegs 234, 239, 253, 255, 257,<br />
258, 259, 260, 264, 265, 266, 267,<br />
269, 270, 272, 275, 277, 278, 281,<br />
282<br />
thin plate spline surfaces 231, 236,<br />
239, 240, 243<br />
tie points 119, 124, 125, 126, 127,<br />
130, 131, 132, 134, 135, 136, 140,<br />
144, 156, 196<br />
time delay integration (TDI) 24, 60<br />
topographic surfaces 166, 167, 229,<br />
231, 232, 233, 235, 236, 239, 243,<br />
254, 257, 258, 266, 281<br />
trajectography 38, 42, 116, 118, 119,<br />
122<br />
triangular irregular networks (TIN)<br />
160, 162–4, 168, 196, 225–6,<br />
231–2, 235, 251<br />
trichromatic vision 26<br />
vignettage 23, 45<br />
visibility 19, 20, 21, 73, 122, 198,<br />
289<br />
volumetric models 326<br />
VRML 325, 333<br />
watersheds 255, 257<br />
wavelet transforms 105, 108, 110,<br />
111, 112, 113<br />
wireframe models 326, 327, 328